lustre/lfsck/lfsck_layout.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2014, 2017, Intel Corporation.
  24  */
  25 /*
  26  * lustre/lfsck/lfsck_layout.c
  27  *
  28  * Author: Fan, Yong <fan.yong@intel.com>
  29  */
  30
  31 #ifndef EXPORT_SYMTAB
  32 # define EXPORT_SYMTAB
  33 #endif
  34 #define DEBUG_SUBSYSTEM S_LFSCK
  35
  36 #include <linux/bitops.h>
  37 #include <linux/rbtree.h>
  38
  39 #include <lu_object.h>
  40 #include <dt_object.h>
  41 #include <lustre_fid.h>
  42 #include <lustre_lib.h>
  43 #include <lustre_net.h>
  44 #include <md_object.h>
  45 #include <obd_class.h>
  46
  47 #include "lfsck_internal.h"
  48
  49 #define LFSCK_LAYOUT_MAGIC_V1           0xB173AE14
  50 #define LFSCK_LAYOUT_MAGIC_V2           0xB1734D76
  51 #define LFSCK_LAYOUT_MAGIC_V3           0xB17371B9
  52 #define LFSCK_LAYOUT_MAGIC_V4           0xB1732FED
  53
  54 #define LFSCK_LAYOUT_MAGIC              LFSCK_LAYOUT_MAGIC_V4
  55
  56 struct lfsck_layout_seq {
  57         struct list_head         lls_list;
  58         __u64                    lls_seq;
  59         __u64                    lls_lastid;
  60         __u64                    lls_lastid_known;
  61         struct dt_object        *lls_lastid_obj;
  62         unsigned int             lls_dirty:1;
  63 };
  64
  65 struct lfsck_layout_slave_target {
  66         /* link into lfsck_layout_slave_data::llsd_master_list. */
  67         struct list_head        llst_list;
  68         /* The position for next record in the rbtree for iteration. */
  69         struct lu_fid           llst_fid;
  70         /* Dummy hash for iteration against the rbtree. */
  71         __u64                   llst_hash;
  72         __u64                   llst_gen;
  73         atomic_t                llst_ref;
  74         __u32                   llst_index;
  75         /* How many times we have failed to get the master status. */
  76         int                     llst_failures;
  77 };
  78
  79 struct lfsck_layout_slave_data {
  80         /* list for lfsck_layout_seq */
  81         struct list_head         llsd_seq_list;
  82
  83         /* list for the masters involve layout verification. */
  84         struct list_head         llsd_master_list;
  85         spinlock_t               llsd_lock;
  86         __u64                    llsd_touch_gen;
  87         struct dt_object        *llsd_rb_obj;
  88         struct rb_root           llsd_rb_root;
  89         struct rw_semaphore      llsd_rb_rwsem;
  90         unsigned int             llsd_rbtree_valid:1;
  91 };
  92
  93 struct lfsck_layout_slave_async_args {
  94         struct obd_export                *llsaa_exp;
  95         struct lfsck_component           *llsaa_com;
  96         struct lfsck_layout_slave_target *llsaa_llst;
  97 };
  98
  99 static inline bool lfsck_comp_extent_aligned(__u64 size)
 100 {
 101          return (size & (LOV_MIN_STRIPE_SIZE - 1)) == 0;
 102 }
 103
 104 static inline void
 105 lfsck_layout_llst_put(struct lfsck_layout_slave_target *llst)
 106 {
 107         if (atomic_dec_and_test(&llst->llst_ref)) {
 108                 LASSERT(list_empty(&llst->llst_list));
 109
 110                 OBD_FREE_PTR(llst);
 111         }
 112 }
 113
 114 static inline int
 115 lfsck_layout_llst_add(struct lfsck_layout_slave_data *llsd, __u32 index)
 116 {
 117         struct lfsck_layout_slave_target *llst;
 118         struct lfsck_layout_slave_target *tmp;
 119         int                               rc   = 0;
 120
 121         OBD_ALLOC_PTR(llst);
 122         if (llst == NULL)
 123                 return -ENOMEM;
 124
 125         INIT_LIST_HEAD(&llst->llst_list);
 126         llst->llst_gen = 0;
 127         llst->llst_index = index;
 128         atomic_set(&llst->llst_ref, 1);
 129
 130         spin_lock(&llsd->llsd_lock);
 131         list_for_each_entry(tmp, &llsd->llsd_master_list, llst_list) {
 132                 if (tmp->llst_index == index) {
 133                         rc = -EALREADY;
 134                         break;
 135                 }
 136         }
 137         if (rc == 0)
 138                 list_add_tail(&llst->llst_list, &llsd->llsd_master_list);
 139         spin_unlock(&llsd->llsd_lock);
 140
 141         if (rc != 0)
 142                 OBD_FREE_PTR(llst);
 143
 144         return rc;
 145 }
 146
 147 static inline void
 148 lfsck_layout_llst_del(struct lfsck_layout_slave_data *llsd,
 149                       struct lfsck_layout_slave_target *llst)
 150 {
 151         bool del = false;
 152
 153         spin_lock(&llsd->llsd_lock);
 154         if (!list_empty(&llst->llst_list)) {
 155                 list_del_init(&llst->llst_list);
 156                 del = true;
 157         }
 158         spin_unlock(&llsd->llsd_lock);
 159
 160         if (del)
 161                 lfsck_layout_llst_put(llst);
 162 }
 163
 164 static inline struct lfsck_layout_slave_target *
 165 lfsck_layout_llst_find_and_del(struct lfsck_layout_slave_data *llsd,
 166                                __u32 index, bool unlink)
 167 {
 168         struct lfsck_layout_slave_target *llst;
 169
 170         spin_lock(&llsd->llsd_lock);
 171         list_for_each_entry(llst, &llsd->llsd_master_list, llst_list) {
 172                 if (llst->llst_index == index) {
 173                         if (unlink)
 174                                 list_del_init(&llst->llst_list);
 175                         else
 176                                 atomic_inc(&llst->llst_ref);
 177                         spin_unlock(&llsd->llsd_lock);
 178
 179                         return llst;
 180                 }
 181         }
 182         spin_unlock(&llsd->llsd_lock);
 183
 184         return NULL;
 185 }
 186
 187 static struct lfsck_layout_req *
 188 lfsck_layout_assistant_req_init(struct lfsck_assistant_object *lso,
 189                                 struct dt_object *child, __u32 comp_id,
 190                                 __u32 ost_idx, __u32 lov_idx)
 191 {
 192         struct lfsck_layout_req *llr;
 193
 194         OBD_ALLOC_PTR(llr);
 195         if (llr == NULL)
 196                 return ERR_PTR(-ENOMEM);
 197
 198         INIT_LIST_HEAD(&llr->llr_lar.lar_list);
 199         llr->llr_lar.lar_parent = lfsck_assistant_object_get(lso);
 200         llr->llr_child = child;
 201         llr->llr_comp_id = comp_id;
 202         llr->llr_ost_idx = ost_idx;
 203         llr->llr_lov_idx = lov_idx;
 204
 205         return llr;
 206 }
 207
 208 static void lfsck_layout_assistant_req_fini(const struct lu_env *env,
 209                                             struct lfsck_assistant_req *lar)
 210 {
 211         struct lfsck_layout_req *llr =
 212                         container_of0(lar, struct lfsck_layout_req, llr_lar);
 213
 214         lfsck_object_put(env, llr->llr_child);
 215         lfsck_assistant_object_put(env, lar->lar_parent);
 216         OBD_FREE_PTR(llr);
 217 }
 218
 219 static int
 220 lfsck_layout_assistant_sync_failures_interpret(const struct lu_env *env,
 221                                                struct ptlrpc_request *req,
 222                                                void *args, int rc)
 223 {
 224         if (rc == 0) {
 225                 struct lfsck_async_interpret_args *laia = args;
 226                 struct lfsck_tgt_desc             *ltd  = laia->laia_ltd;
 227
 228                 ltd->ltd_synced_failures = 1;
 229                 atomic_dec(laia->laia_count);
 230         }
 231
 232         return 0;
 233 }
 234
 235 /**
 236  * Notify remote LFSCK instances about former failures.
 237  *
 238  * The local LFSCK instance has recorded which OSTs have ever failed to respond
 239  * some LFSCK verification requests (maybe because of network issues or the OST
 240  * itself trouble). During the respond gap, the OST may missed some OST-objects
 241  * verification, then the OST cannot know whether related OST-objects have been
 242  * referenced by related MDT-objects or not, then in the second-stage scanning,
 243  * these OST-objects will be regarded as orphan, if the OST-object contains bad
 244  * parent FID for back reference, then it will misguide the LFSCK to make wrong
 245  * fixing for the fake orphan.
 246  *
 247  * To avoid above trouble, when layout LFSCK finishes the first-stage scanning,
 248  * it will scan the bitmap for the ever failed OSTs, and notify them that they
 249  * have ever missed some OST-object verification and should skip the handling
 250  * for orphan OST-objects on all MDTs that are in the layout LFSCK.
 251  *
 252  * \param[in] env       pointer to the thread context
 253  * \param[in] com       pointer to the lfsck component
 254  * \param[in] lr        pointer to the lfsck request
 255  */
 256 static void lfsck_layout_assistant_sync_failures(const struct lu_env *env,
 257                                                  struct lfsck_component *com,
 258                                                  struct lfsck_request *lr)
 259 {
 260         struct lfsck_async_interpret_args *laia  =
 261                                 &lfsck_env_info(env)->lti_laia2;
 262         struct lfsck_assistant_data       *lad   = com->lc_data;
 263         struct lfsck_layout               *lo    = com->lc_file_ram;
 264         struct lfsck_instance             *lfsck = com->lc_lfsck;
 265         struct lfsck_tgt_descs            *ltds  = &lfsck->li_ost_descs;
 266         struct lfsck_tgt_desc             *ltd;
 267         struct ptlrpc_request_set         *set;
 268         atomic_t                           count;
 269         __u32                              idx;
 270         int                                rc    = 0;
 271         ENTRY;
 272
 273         if (!test_bit(LAD_INCOMPLETE, &lad->lad_flags))
 274                 RETURN_EXIT;
 275
 276         /* If the MDT has ever failed to verfiy some OST-objects,
 277          * then sync failures with them firstly. */
 278         lr->lr_flags2 = lo->ll_flags | LF_INCOMPLETE;
 279
 280         atomic_set(&count, 0);
 281         memset(laia, 0, sizeof(*laia));
 282         laia->laia_count = &count;
 283         set = ptlrpc_prep_set();
 284         if (set == NULL)
 285                 GOTO(out, rc = -ENOMEM);
 286
 287         down_read(&ltds->ltd_rw_sem);
 288         cfs_foreach_bit(lad->lad_bitmap, idx) {
 289                 ltd = lfsck_ltd2tgt(ltds, idx);
 290                 if (unlikely(!ltd))
 291                         continue;
 292
 293                 laia->laia_ltd = ltd;
 294                 rc = lfsck_async_request(env, ltd->ltd_exp, lr, set,
 295                                 lfsck_layout_assistant_sync_failures_interpret,
 296                                 laia, LFSCK_NOTIFY);
 297                 if (rc != 0) {
 298                         CDEBUG(D_LFSCK, "%s: LFSCK assistant fail to "
 299                                "notify target %x for %s phase1 done: "
 300                                "rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
 301                                ltd->ltd_index, lad->lad_name, rc);
 302
 303                         break;
 304                 }
 305
 306                 atomic_inc(&count);
 307         }
 308         up_read(&ltds->ltd_rw_sem);
 309
 310         if (rc == 0 && atomic_read(&count) > 0)
 311                 rc = ptlrpc_set_wait(env, set);
 312
 313         ptlrpc_set_destroy(set);
 314
 315         if (rc == 0 && atomic_read(&count) > 0)
 316                 rc = -EINVAL;
 317
 318         GOTO(out, rc);
 319
 320 out:
 321         if (rc != 0)
 322                 /* If failed to sync failures with the OSTs, then have to
 323                  * mark the whole LFSCK as LF_INCOMPLETE to skip the whole
 324                  * subsequent orphan OST-object handling. */
 325                 lo->ll_flags |= LF_INCOMPLETE;
 326
 327         lr->lr_flags2 = lo->ll_flags;
 328 }
 329
 330 static int lfsck_layout_verify_header_v1v3(struct dt_object *obj,
 331                                            struct lov_mds_md_v1 *lmm,
 332                                            __u64 start, __u32 comp_id)
 333 {
 334         __u32 magic;
 335         __u32 pattern;
 336
 337         magic = le32_to_cpu(lmm->lmm_magic);
 338         /* If magic crashed, keep it there. Sometime later, during OST-object
 339          * orphan handling, if some OST-object(s) back-point to it, it can be
 340          * verified and repaired. */
 341         if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3) {
 342                 int rc;
 343
 344                 if ((magic & LOV_MAGIC_MASK) == LOV_MAGIC_MAGIC)
 345                         rc = -EOPNOTSUPP;
 346                 else
 347                         rc = -EINVAL;
 348
 349                 CDEBUG(D_LFSCK, "%s LOV EA magic %u for the file "DFID"\n",
 350                        rc == -EINVAL ? "Unknown" : "Unsupported",
 351                        magic, PFID(lfsck_dto2fid(obj)));
 352
 353                 return rc;
 354         }
 355
 356         pattern = le32_to_cpu(lmm->lmm_pattern);
 357
 358 #if 0
 359         /* XXX: DoM file verification will be supportted via LU-11081. */
 360         if (lov_pattern(pattern) == LOV_PATTERN_MDT) {
 361                 if (start != 0) {
 362                         CDEBUG(D_LFSCK, "The DoM entry for "DFID" is not "
 363                                "the first component in the mirror %x/%llu\n",
 364                                PFID(lfsck_dto2fid(obj)), comp_id, start);
 365
 366                         return -EINVAL;
 367                 }
 368         }
 369 #endif
 370
 371         if (!lov_pattern_supported_normal_comp(lov_pattern(pattern))) {
 372                 CDEBUG(D_LFSCK, "Unsupported LOV EA pattern %u for the file "
 373                        DFID" in the component %x\n",
 374                        pattern, PFID(lfsck_dto2fid(obj)), comp_id);
 375
 376                 return -EOPNOTSUPP;
 377         }
 378
 379         return 0;
 380 }
 381
 382 static int lfsck_layout_verify_header_foreign(struct dt_object *obj,
 383                                               struct lov_foreign_md *lfm,
 384                                               size_t len)
 385 {
 386         /* magic has been verified already */
 387         __u32 value_len = le32_to_cpu(lfm->lfm_length);
 388         /* type and flags are not checked for instance */
 389
 390         CDEBUG(D_INFO, "foreign LOV EA, magic %x, len %u, type %x, flags %x, for file "DFID"\n",
 391                le32_to_cpu(lfm->lfm_magic), value_len,
 392                le32_to_cpu(lfm->lfm_type), le32_to_cpu(lfm->lfm_flags),
 393                PFID(lfsck_dto2fid(obj)));
 394
 395         if (len != value_len + offsetof(typeof(*lfm), lfm_value))
 396                 CDEBUG(D_LFSCK, "foreign LOV EA internal size %u does not match EA full size %zu for file "DFID"\n",
 397                        value_len, len, PFID(lfsck_dto2fid(obj)));
 398
 399         /* nothing to repair */
 400         return -ENODATA;
 401 }
 402
 403 static int lfsck_layout_verify_header(struct dt_object *obj,
 404                                       struct lov_mds_md_v1 *lmm, size_t len)
 405 {
 406         int rc = 0;
 407
 408         if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
 409                 struct lov_comp_md_v1 *lcm = (struct lov_comp_md_v1 *)lmm;
 410                 int i;
 411                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
 412
 413                 if (unlikely(count == 0)) {
 414                         CDEBUG(D_LFSCK, "the PFL file "DFID" contains invalid "
 415                                "components count 0\n",
 416                                PFID(lfsck_dto2fid(obj)));
 417
 418                         return -EINVAL;
 419                 }
 420
 421                 for (i = 0; i < count && !rc; i++) {
 422                         struct lov_comp_md_entry_v1 *lcme =
 423                                                 &lcm->lcm_entries[i];
 424                         __u64 start = le64_to_cpu(lcme->lcme_extent.e_start);
 425                         __u64 end = le64_to_cpu(lcme->lcme_extent.e_end);
 426                         __u32 comp_id = le32_to_cpu(lcme->lcme_id);
 427
 428                         if (unlikely(comp_id == LCME_ID_INVAL ||
 429                                      comp_id > LCME_ID_MAX)) {
 430                                 CDEBUG(D_LFSCK, "found invalid FPL ID %u "
 431                                        "for the file "DFID" at idx %d\n",
 432                                        comp_id, PFID(lfsck_dto2fid(obj)), i);
 433
 434                                 return -EINVAL;
 435                         }
 436
 437                         if (unlikely(start >= end ||
 438                                      !lfsck_comp_extent_aligned(start) ||
 439                                      (!lfsck_comp_extent_aligned(end) &&
 440                                       end != LUSTRE_EOF))) {
 441                                 CDEBUG(D_LFSCK, "found invalid FPL extent "
 442                                        "range [%llu - %llu) for the file "
 443                                        DFID" at idx %d\n",
 444                                        start, end, PFID(lfsck_dto2fid(obj)), i);
 445
 446                                 return -EINVAL;
 447                         }
 448
 449                         rc = lfsck_layout_verify_header_v1v3(obj,
 450                                         (struct lov_mds_md_v1 *)((char *)lmm +
 451                                         le32_to_cpu(lcme->lcme_offset)), start,
 452                                         comp_id);
 453                 }
 454         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_FOREIGN) {
 455                 rc = lfsck_layout_verify_header_foreign(obj,
 456                                                 (struct lov_foreign_md *)lmm,
 457                                                 len);
 458         } else {
 459                 rc = lfsck_layout_verify_header_v1v3(obj, lmm, 1, 0);
 460         }
 461
 462         return rc;
 463 }
 464
 465 static int lfsck_layout_get_lovea(const struct lu_env *env,
 466                                   struct dt_object *obj, struct lu_buf *buf)
 467 {
 468         int rc;
 469         int rc1;
 470
 471 again:
 472         rc = dt_xattr_get(env, obj, buf, XATTR_NAME_LOV);
 473         if (rc == -ERANGE) {
 474                 rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_LOV);
 475                 if (rc <= 0)
 476                         return !rc ? -ENODATA : rc;
 477
 478                 lu_buf_realloc(buf, rc);
 479                 if (buf->lb_buf == NULL)
 480                         return -ENOMEM;
 481
 482                 goto again;
 483         }
 484
 485         if (rc <= 0)
 486                 return !rc ? -ENODATA : rc;
 487
 488         if (unlikely(buf->lb_buf == NULL)) {
 489                 lu_buf_alloc(buf, rc);
 490                 if (buf->lb_buf == NULL)
 491                         return -ENOMEM;
 492
 493                 goto again;
 494         }
 495
 496         rc1 = lfsck_layout_verify_header(obj, buf->lb_buf, rc);
 497
 498         return rc1 ? rc1 : rc;
 499 }
 500
 501 #define LFSCK_RBTREE_BITMAP_SIZE        PAGE_SIZE
 502 #define LFSCK_RBTREE_BITMAP_WIDTH       (LFSCK_RBTREE_BITMAP_SIZE << 3)
 503 #define LFSCK_RBTREE_BITMAP_MASK        (LFSCK_RBTREE_BITMAP_WIDTH - 1)
 504
 505 struct lfsck_rbtree_node {
 506         struct rb_node   lrn_node;
 507         __u64            lrn_seq;
 508         __u32            lrn_first_oid;
 509         atomic_t         lrn_known_count;
 510         atomic_t         lrn_accessed_count;
 511         void            *lrn_known_bitmap;
 512         void            *lrn_accessed_bitmap;
 513 };
 514
 515 static inline int lfsck_rbtree_cmp(struct lfsck_rbtree_node *lrn,
 516                                    __u64 seq, __u32 oid)
 517 {
 518         if (seq < lrn->lrn_seq)
 519                 return -1;
 520
 521         if (seq > lrn->lrn_seq)
 522                 return 1;
 523
 524         if (oid < lrn->lrn_first_oid)
 525                 return -1;
 526
 527         if (oid - lrn->lrn_first_oid >= LFSCK_RBTREE_BITMAP_WIDTH)
 528                 return 1;
 529
 530         return 0;
 531 }
 532
 533 /* The caller should hold llsd->llsd_rb_lock. */
 534 static struct lfsck_rbtree_node *
 535 lfsck_rbtree_search(struct lfsck_layout_slave_data *llsd,
 536                     const struct lu_fid *fid, bool *exact)
 537 {
 538         struct rb_node           *node  = llsd->llsd_rb_root.rb_node;
 539         struct rb_node           *prev  = NULL;
 540         struct lfsck_rbtree_node *lrn   = NULL;
 541         int                       rc    = 0;
 542
 543         if (exact != NULL)
 544                 *exact = true;
 545
 546         while (node != NULL) {
 547                 prev = node;
 548                 lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 549                 rc = lfsck_rbtree_cmp(lrn, fid_seq(fid), fid_oid(fid));
 550                 if (rc < 0)
 551                         node = node->rb_left;
 552                 else if (rc > 0)
 553                         node = node->rb_right;
 554                 else
 555                         return lrn;
 556         }
 557
 558         if (exact == NULL)
 559                 return NULL;
 560
 561         /* If there is no exactly matched one, then to the next valid one. */
 562         *exact = false;
 563
 564         /* The rbtree is empty. */
 565         if (rc == 0)
 566                 return NULL;
 567
 568         if (rc < 0)
 569                 return lrn;
 570
 571         node = rb_next(prev);
 572
 573         /* The end of the rbtree. */
 574         if (node == NULL)
 575                 return NULL;
 576
 577         lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 578
 579         return lrn;
 580 }
 581
 582 static struct lfsck_rbtree_node *lfsck_rbtree_new(const struct lu_env *env,
 583                                                   const struct lu_fid *fid)
 584 {
 585         struct lfsck_rbtree_node *lrn;
 586
 587         OBD_ALLOC_PTR(lrn);
 588         if (lrn == NULL)
 589                 return ERR_PTR(-ENOMEM);
 590
 591         OBD_ALLOC(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 592         if (lrn->lrn_known_bitmap == NULL) {
 593                 OBD_FREE_PTR(lrn);
 594
 595                 return ERR_PTR(-ENOMEM);
 596         }
 597
 598         OBD_ALLOC(lrn->lrn_accessed_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 599         if (lrn->lrn_accessed_bitmap == NULL) {
 600                 OBD_FREE(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 601                 OBD_FREE_PTR(lrn);
 602
 603                 return ERR_PTR(-ENOMEM);
 604         }
 605
 606         RB_CLEAR_NODE(&lrn->lrn_node);
 607         lrn->lrn_seq = fid_seq(fid);
 608         lrn->lrn_first_oid = fid_oid(fid) & ~LFSCK_RBTREE_BITMAP_MASK;
 609         atomic_set(&lrn->lrn_known_count, 0);
 610         atomic_set(&lrn->lrn_accessed_count, 0);
 611
 612         return lrn;
 613 }
 614
 615 static void lfsck_rbtree_free(struct lfsck_rbtree_node *lrn)
 616 {
 617         OBD_FREE(lrn->lrn_accessed_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 618         OBD_FREE(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 619         OBD_FREE_PTR(lrn);
 620 }
 621
 622 /* The caller should hold lock. */
 623 static struct lfsck_rbtree_node *
 624 lfsck_rbtree_insert(struct lfsck_layout_slave_data *llsd,
 625                     struct lfsck_rbtree_node *lrn)
 626 {
 627         struct rb_node           **pos    = &llsd->llsd_rb_root.rb_node;
 628         struct rb_node            *parent = NULL;
 629         struct lfsck_rbtree_node  *tmp;
 630         int                        rc;
 631
 632         while (*pos != NULL) {
 633                 parent = *pos;
 634                 tmp = rb_entry(parent, struct lfsck_rbtree_node, lrn_node);
 635                 rc = lfsck_rbtree_cmp(tmp, lrn->lrn_seq, lrn->lrn_first_oid);
 636                 if (rc < 0)
 637                         pos = &(*pos)->rb_left;
 638                 else if (rc > 0)
 639                         pos = &(*pos)->rb_right;
 640                 else
 641                         return tmp;
 642         }
 643
 644         rb_link_node(&lrn->lrn_node, parent, pos);
 645         rb_insert_color(&lrn->lrn_node, &llsd->llsd_rb_root);
 646
 647         return lrn;
 648 }
 649
 650 static const struct dt_index_operations lfsck_orphan_index_ops;
 651
 652 static int lfsck_rbtree_setup(const struct lu_env *env,
 653                               struct lfsck_component *com)
 654 {
 655         struct lu_fid                   *fid    = &lfsck_env_info(env)->lti_fid;
 656         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 657         struct dt_device                *dev    = lfsck->li_bottom;
 658         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
 659         struct dt_object                *obj;
 660
 661         fid->f_seq = FID_SEQ_LAYOUT_RBTREE;
 662         fid->f_oid = lfsck_dev_idx(lfsck);
 663         fid->f_ver = 0;
 664         obj = dt_locate(env, dev, fid);
 665         if (IS_ERR(obj))
 666                 RETURN(PTR_ERR(obj));
 667
 668         /* Generate an in-RAM object to stand for the layout rbtree.
 669          * Scanning the layout rbtree will be via the iteration over
 670          * the object. In the future, the rbtree may be written onto
 671          * disk with the object.
 672          *
 673          * Mark the object to be as exist. */
 674         obj->do_lu.lo_header->loh_attr |= LOHA_EXISTS;
 675         obj->do_index_ops = &lfsck_orphan_index_ops;
 676         llsd->llsd_rb_obj = obj;
 677         llsd->llsd_rbtree_valid = 1;
 678         dev->dd_record_fid_accessed = 1;
 679
 680         CDEBUG(D_LFSCK, "%s: layout LFSCK init OST-objects accessing bitmap\n",
 681                lfsck_lfsck2name(lfsck));
 682
 683         return 0;
 684 }
 685
 686 static void lfsck_rbtree_cleanup(const struct lu_env *env,
 687                                  struct lfsck_component *com)
 688 {
 689         struct lfsck_instance           *lfsck = com->lc_lfsck;
 690         struct lfsck_layout_slave_data  *llsd  = com->lc_data;
 691         struct rb_node                  *node  = rb_first(&llsd->llsd_rb_root);
 692         struct rb_node                  *next;
 693         struct lfsck_rbtree_node        *lrn;
 694
 695         lfsck->li_bottom->dd_record_fid_accessed = 0;
 696         /* Invalid the rbtree, then no others will use it. */
 697         down_write(&llsd->llsd_rb_rwsem);
 698         llsd->llsd_rbtree_valid = 0;
 699         up_write(&llsd->llsd_rb_rwsem);
 700
 701         while (node != NULL) {
 702                 next = rb_next(node);
 703                 lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 704                 rb_erase(node, &llsd->llsd_rb_root);
 705                 lfsck_rbtree_free(lrn);
 706                 node = next;
 707         }
 708
 709         if (llsd->llsd_rb_obj != NULL) {
 710                 lfsck_object_put(env, llsd->llsd_rb_obj);
 711                 llsd->llsd_rb_obj = NULL;
 712         }
 713
 714         CDEBUG(D_LFSCK, "%s: layout LFSCK fini OST-objects accessing bitmap\n",
 715                lfsck_lfsck2name(lfsck));
 716 }
 717
 718 static void lfsck_rbtree_update_bitmap(const struct lu_env *env,
 719                                        struct lfsck_component *com,
 720                                        const struct lu_fid *fid,
 721                                        bool accessed)
 722 {
 723         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
 724         struct lfsck_rbtree_node        *lrn;
 725         bool                             insert = false;
 726         int                              idx;
 727         int                              rc     = 0;
 728         ENTRY;
 729
 730         if (unlikely(!fid_is_sane(fid) || fid_is_last_id(fid)))
 731                 RETURN_EXIT;
 732
 733         if (!fid_is_idif(fid) && !fid_is_norm(fid))
 734                 RETURN_EXIT;
 735
 736         down_read(&llsd->llsd_rb_rwsem);
 737         if (!llsd->llsd_rbtree_valid)
 738                 GOTO(unlock, rc = 0);
 739
 740         lrn = lfsck_rbtree_search(llsd, fid, NULL);
 741         if (lrn == NULL) {
 742                 struct lfsck_rbtree_node *tmp;
 743
 744                 LASSERT(!insert);
 745
 746                 up_read(&llsd->llsd_rb_rwsem);
 747                 tmp = lfsck_rbtree_new(env, fid);
 748                 if (IS_ERR(tmp))
 749                         GOTO(out, rc = PTR_ERR(tmp));
 750
 751                 insert = true;
 752                 down_write(&llsd->llsd_rb_rwsem);
 753                 if (!llsd->llsd_rbtree_valid) {
 754                         lfsck_rbtree_free(tmp);
 755                         GOTO(unlock, rc = 0);
 756                 }
 757
 758                 lrn = lfsck_rbtree_insert(llsd, tmp);
 759                 if (lrn != tmp)
 760                         lfsck_rbtree_free(tmp);
 761         }
 762
 763         idx = fid_oid(fid) & LFSCK_RBTREE_BITMAP_MASK;
 764         /* Any accessed object must be a known object. */
 765         if (!test_and_set_bit(idx, lrn->lrn_known_bitmap))
 766                 atomic_inc(&lrn->lrn_known_count);
 767         if (accessed && !test_and_set_bit(idx, lrn->lrn_accessed_bitmap))
 768                 atomic_inc(&lrn->lrn_accessed_count);
 769
 770         GOTO(unlock, rc = 0);
 771
 772 unlock:
 773         if (insert)
 774                 up_write(&llsd->llsd_rb_rwsem);
 775         else
 776                 up_read(&llsd->llsd_rb_rwsem);
 777 out:
 778         if (rc != 0 && accessed) {
 779                 struct lfsck_layout *lo = com->lc_file_ram;
 780
 781                 CDEBUG(D_LFSCK, "%s: fail to update OST-objects accessing "
 782                        "bitmap, and will cause incorrect LFSCK OST-object "
 783                        "handling, so disable it to cancel orphan handling "
 784                        "for related device. rc = %d\n",
 785                        lfsck_lfsck2name(com->lc_lfsck), rc);
 786
 787                 lo->ll_flags |= LF_INCOMPLETE;
 788                 lfsck_rbtree_cleanup(env, com);
 789         }
 790 }
 791
 792 static inline void lldk_le_to_cpu(struct lfsck_layout_dangling_key *des,
 793                                   const struct lfsck_layout_dangling_key *src)
 794 {
 795         fid_le_to_cpu(&des->lldk_fid, &src->lldk_fid);
 796         des->lldk_comp_id = le32_to_cpu(src->lldk_comp_id);
 797         des->lldk_ea_off = le32_to_cpu(src->lldk_ea_off);
 798 }
 799
 800 static inline void lldk_cpu_to_le(struct lfsck_layout_dangling_key *des,
 801                                   const struct lfsck_layout_dangling_key *src)
 802 {
 803         fid_cpu_to_le(&des->lldk_fid, &src->lldk_fid);
 804         des->lldk_comp_id = cpu_to_le32(src->lldk_comp_id);
 805         des->lldk_ea_off = cpu_to_le32(src->lldk_ea_off);
 806 }
 807
 808 static inline void lldk_be_to_cpu(struct lfsck_layout_dangling_key *des,
 809                                   const struct lfsck_layout_dangling_key *src)
 810 {
 811         fid_be_to_cpu(&des->lldk_fid, &src->lldk_fid);
 812         des->lldk_comp_id = be32_to_cpu(src->lldk_comp_id);
 813         des->lldk_ea_off = be32_to_cpu(src->lldk_ea_off);
 814 }
 815
 816 static inline void lldk_cpu_to_be(struct lfsck_layout_dangling_key *des,
 817                                   const struct lfsck_layout_dangling_key *src)
 818 {
 819         fid_cpu_to_be(&des->lldk_fid, &src->lldk_fid);
 820         des->lldk_comp_id = cpu_to_be32(src->lldk_comp_id);
 821         des->lldk_ea_off = cpu_to_be32(src->lldk_ea_off);
 822 }
 823
 824 static void lfsck_layout_le_to_cpu(struct lfsck_layout *des,
 825                                    const struct lfsck_layout *src)
 826 {
 827         int i;
 828
 829         des->ll_magic = le32_to_cpu(src->ll_magic);
 830         des->ll_status = le32_to_cpu(src->ll_status);
 831         des->ll_flags = le32_to_cpu(src->ll_flags);
 832         des->ll_success_count = le32_to_cpu(src->ll_success_count);
 833         des->ll_run_time_phase1 = le64_to_cpu(src->ll_run_time_phase1);
 834         des->ll_run_time_phase2 = le64_to_cpu(src->ll_run_time_phase2);
 835         des->ll_time_last_complete = le64_to_cpu(src->ll_time_last_complete);
 836         des->ll_time_latest_start = le64_to_cpu(src->ll_time_latest_start);
 837         des->ll_time_last_checkpoint =
 838                                 le64_to_cpu(src->ll_time_last_checkpoint);
 839         des->ll_pos_latest_start = le64_to_cpu(src->ll_pos_latest_start);
 840         des->ll_pos_last_checkpoint = le64_to_cpu(src->ll_pos_last_checkpoint);
 841         des->ll_pos_first_inconsistent =
 842                         le64_to_cpu(src->ll_pos_first_inconsistent);
 843         des->ll_objs_checked_phase1 = le64_to_cpu(src->ll_objs_checked_phase1);
 844         des->ll_objs_failed_phase1 = le64_to_cpu(src->ll_objs_failed_phase1);
 845         des->ll_objs_checked_phase2 = le64_to_cpu(src->ll_objs_checked_phase2);
 846         des->ll_objs_failed_phase2 = le64_to_cpu(src->ll_objs_failed_phase2);
 847         for (i = 0; i < LLIT_MAX; i++)
 848                 des->ll_objs_repaired[i] =
 849                                 le64_to_cpu(src->ll_objs_repaired[i]);
 850         des->ll_objs_skipped = le64_to_cpu(src->ll_objs_skipped);
 851         des->ll_bitmap_size = le32_to_cpu(src->ll_bitmap_size);
 852         lldk_le_to_cpu(&des->ll_lldk_latest_scanned_phase2,
 853                        &src->ll_lldk_latest_scanned_phase2);
 854 }
 855
 856 static void lfsck_layout_cpu_to_le(struct lfsck_layout *des,
 857                                    const struct lfsck_layout *src)
 858 {
 859         int i;
 860
 861         des->ll_magic = cpu_to_le32(src->ll_magic);
 862         des->ll_status = cpu_to_le32(src->ll_status);
 863         des->ll_flags = cpu_to_le32(src->ll_flags);
 864         des->ll_success_count = cpu_to_le32(src->ll_success_count);
 865         des->ll_run_time_phase1 = cpu_to_le64(src->ll_run_time_phase1);
 866         des->ll_run_time_phase2 = cpu_to_le64(src->ll_run_time_phase2);
 867         des->ll_time_last_complete = cpu_to_le64(src->ll_time_last_complete);
 868         des->ll_time_latest_start = cpu_to_le64(src->ll_time_latest_start);
 869         des->ll_time_last_checkpoint =
 870                                 cpu_to_le64(src->ll_time_last_checkpoint);
 871         des->ll_pos_latest_start = cpu_to_le64(src->ll_pos_latest_start);
 872         des->ll_pos_last_checkpoint = cpu_to_le64(src->ll_pos_last_checkpoint);
 873         des->ll_pos_first_inconsistent =
 874                         cpu_to_le64(src->ll_pos_first_inconsistent);
 875         des->ll_objs_checked_phase1 = cpu_to_le64(src->ll_objs_checked_phase1);
 876         des->ll_objs_failed_phase1 = cpu_to_le64(src->ll_objs_failed_phase1);
 877         des->ll_objs_checked_phase2 = cpu_to_le64(src->ll_objs_checked_phase2);
 878         des->ll_objs_failed_phase2 = cpu_to_le64(src->ll_objs_failed_phase2);
 879         for (i = 0; i < LLIT_MAX; i++)
 880                 des->ll_objs_repaired[i] =
 881                                 cpu_to_le64(src->ll_objs_repaired[i]);
 882         des->ll_objs_skipped = cpu_to_le64(src->ll_objs_skipped);
 883         des->ll_bitmap_size = cpu_to_le32(src->ll_bitmap_size);
 884         lldk_cpu_to_le(&des->ll_lldk_latest_scanned_phase2,
 885                        &src->ll_lldk_latest_scanned_phase2);
 886 }
 887
 888 /**
 889  * Load the OST bitmap from the lfsck_layout trace file.
 890  *
 891  * \param[in] env       pointer to the thread context
 892  * \param[in] com       pointer to the lfsck component
 893  *
 894  * \retval              0 for success
 895  * \retval              negative error number on failure or data corruption
 896  */
 897 static int lfsck_layout_load_bitmap(const struct lu_env *env,
 898                                     struct lfsck_component *com)
 899 {
 900         struct dt_object                *obj    = com->lc_obj;
 901         struct lfsck_assistant_data     *lad    = com->lc_data;
 902         struct lfsck_layout             *lo     = com->lc_file_ram;
 903         struct cfs_bitmap                       *bitmap = lad->lad_bitmap;
 904         loff_t                           pos    = com->lc_file_size;
 905         ssize_t                          size;
 906         __u32                            nbits;
 907         int                              rc;
 908         ENTRY;
 909
 910         if (com->lc_lfsck->li_ost_descs.ltd_tgts_bitmap->size >
 911             lo->ll_bitmap_size)
 912                 nbits = com->lc_lfsck->li_ost_descs.ltd_tgts_bitmap->size;
 913         else
 914                 nbits = lo->ll_bitmap_size;
 915
 916         if (unlikely(nbits < BITS_PER_LONG))
 917                 nbits = BITS_PER_LONG;
 918
 919         if (nbits > bitmap->size) {
 920                 __u32 new_bits = bitmap->size;
 921                 struct cfs_bitmap *new_bitmap;
 922
 923                 while (new_bits < nbits)
 924                         new_bits <<= 1;
 925
 926                 new_bitmap = CFS_ALLOCATE_BITMAP(new_bits);
 927                 if (new_bitmap == NULL)
 928                         RETURN(-ENOMEM);
 929
 930                 lad->lad_bitmap = new_bitmap;
 931                 CFS_FREE_BITMAP(bitmap);
 932                 bitmap = new_bitmap;
 933         }
 934
 935         if (lo->ll_bitmap_size == 0) {
 936                 clear_bit(LAD_INCOMPLETE, &lad->lad_flags);
 937                 CFS_RESET_BITMAP(bitmap);
 938
 939                 RETURN(0);
 940         }
 941
 942         size = (lo->ll_bitmap_size + 7) >> 3;
 943         rc = dt_read(env, obj, lfsck_buf_get(env, bitmap->data, size), &pos);
 944         if (rc != size)
 945                 RETURN(rc >= 0 ? -EINVAL : rc);
 946
 947         if (cfs_bitmap_check_empty(bitmap))
 948                 clear_bit(LAD_INCOMPLETE, &lad->lad_flags);
 949         else
 950                 set_bit(LAD_INCOMPLETE, &lad->lad_flags);
 951
 952         RETURN(0);
 953 }
 954
 955 /**
 956  * Load the layout LFSCK trace file from disk.
 957  *
 958  * The layout LFSCK trace file records the layout LFSCK status information
 959  * and other statistics, such as how many objects have been scanned, and how
 960  * many objects have been repaired, and etc. It also contains the bitmap for
 961  * failed OSTs during the layout LFSCK. All these information will be loaded
 962  * from disk to RAM when the layout LFSCK component setup.
 963  *
 964  * \param[in] env       pointer to the thread context
 965  * \param[in] com       pointer to the lfsck component
 966  *
 967  * \retval              positive number for file data corruption, the caller
 968  *                      should reset the layout LFSCK trace file
 969  * \retval              0 for success
 970  * \retval              negative error number on failure
 971  */
 972 static int lfsck_layout_load(const struct lu_env *env,
 973                              struct lfsck_component *com)
 974 {
 975         struct lfsck_layout             *lo     = com->lc_file_ram;
 976         ssize_t                          size   = com->lc_file_size;
 977         loff_t                           pos    = 0;
 978         int                              rc;
 979
 980         rc = dt_read(env, com->lc_obj,
 981                      lfsck_buf_get(env, com->lc_file_disk, size), &pos);
 982         if (rc == 0) {
 983                 return -ENOENT;
 984         } else if (rc < 0) {
 985                 CDEBUG(D_LFSCK, "%s: failed to load lfsck_layout: rc = %d\n",
 986                        lfsck_lfsck2name(com->lc_lfsck), rc);
 987                 return rc;
 988         } else if (rc != size) {
 989                 CDEBUG(D_LFSCK, "%s: lfsck_layout size %u != %u; reset it\n",
 990                        lfsck_lfsck2name(com->lc_lfsck), rc, (unsigned int)size);
 991                 return 1;
 992         }
 993
 994         lfsck_layout_le_to_cpu(lo, com->lc_file_disk);
 995         if (lo->ll_magic != LFSCK_LAYOUT_MAGIC) {
 996                 CDEBUG(D_LFSCK, "%s: invalid lfsck_layout magic %#x != %#x, "
 997                        "to be reset\n", lfsck_lfsck2name(com->lc_lfsck),
 998                        lo->ll_magic, LFSCK_LAYOUT_MAGIC);
 999                 return 1;
1000         }
1001
1002         return 0;
1003 }
1004
1005 /**
1006  * Store the layout LFSCK trace file on disk.
1007  *
1008  * The layout LFSCK trace file records the layout LFSCK status information
1009  * and other statistics, such as how many objects have been scanned, and how
1010  * many objects have been repaired, and etc. It also contains the bitmap for
1011  * failed OSTs during the layout LFSCK. All these information will be synced
1012  * from RAM to disk periodically.
1013  *
1014  * \param[in] env       pointer to the thread context
1015  * \param[in] com       pointer to the lfsck component
1016  *
1017  * \retval              0 for success
1018  * \retval              negative error number on failure
1019  */
1020 static int lfsck_layout_store(const struct lu_env *env,
1021                               struct lfsck_component *com)
1022 {
1023         struct dt_object        *obj    = com->lc_obj;
1024         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1025         struct lfsck_layout     *lo_ram = com->lc_file_ram;
1026         struct lfsck_layout     *lo     = com->lc_file_disk;
1027         struct thandle          *th;
1028         struct dt_device        *dev    = lfsck_obj2dev(obj);
1029         struct cfs_bitmap       *bitmap = NULL;
1030         loff_t                   pos;
1031         ssize_t                  size   = com->lc_file_size;
1032         __u32                    nbits  = 0;
1033         int                      rc;
1034         ENTRY;
1035
1036         if (lfsck->li_master) {
1037                 struct lfsck_assistant_data *lad = com->lc_data;
1038
1039                 bitmap = lad->lad_bitmap;
1040                 nbits = bitmap->size;
1041
1042                 LASSERT(nbits > 0);
1043                 LASSERTF((nbits & 7) == 0, "Invalid nbits %u\n", nbits);
1044         }
1045
1046         lo_ram->ll_bitmap_size = nbits;
1047         lfsck_layout_cpu_to_le(lo, lo_ram);
1048         th = dt_trans_create(env, dev);
1049         if (IS_ERR(th))
1050                 GOTO(log, rc = PTR_ERR(th));
1051
1052         rc = dt_declare_record_write(env, obj, lfsck_buf_get(env, lo, size),
1053                                      (loff_t)0, th);
1054         if (rc != 0)
1055                 GOTO(out, rc);
1056
1057         if (bitmap != NULL) {
1058                 rc = dt_declare_record_write(env, obj,
1059                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
1060                                 (loff_t)size, th);
1061                 if (rc != 0)
1062                         GOTO(out, rc);
1063         }
1064
1065         rc = dt_trans_start_local(env, dev, th);
1066         if (rc != 0)
1067                 GOTO(out, rc);
1068
1069         pos = 0;
1070         rc = dt_record_write(env, obj, lfsck_buf_get(env, lo, size), &pos, th);
1071         if (rc != 0)
1072                 GOTO(out, rc);
1073
1074         if (bitmap != NULL) {
1075                 pos = size;
1076                 rc = dt_record_write(env, obj,
1077                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
1078                                 &pos, th);
1079         }
1080
1081         GOTO(out, rc);
1082
1083 out:
1084         dt_trans_stop(env, dev, th);
1085
1086 log:
1087         if (rc != 0)
1088                 CDEBUG(D_LFSCK, "%s: fail to store lfsck_layout: rc = %d\n",
1089                        lfsck_lfsck2name(lfsck), rc);
1090
1091         return rc;
1092 }
1093
1094 static int lfsck_layout_init(const struct lu_env *env,
1095                              struct lfsck_component *com)
1096 {
1097         struct lfsck_layout *lo = com->lc_file_ram;
1098         int rc;
1099
1100         memset(lo, 0, com->lc_file_size);
1101         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
1102         lo->ll_status = LS_INIT;
1103         down_write(&com->lc_sem);
1104         rc = lfsck_layout_store(env, com);
1105         if (rc == 0 && com->lc_lfsck->li_master)
1106                 rc = lfsck_load_sub_trace_files(env, com,
1107                         &dt_lfsck_layout_dangling_features, LFSCK_LAYOUT, true);
1108         up_write(&com->lc_sem);
1109
1110         return rc;
1111 }
1112
1113 static int fid_is_for_ostobj(const struct lu_env *env,
1114                              struct lfsck_instance *lfsck,
1115                              struct dt_object *obj, const struct lu_fid *fid)
1116 {
1117         struct seq_server_site  *ss     = lfsck_dev_site(lfsck);
1118         struct lu_seq_range     *range  = &lfsck_env_info(env)->lti_range;
1119         struct lustre_ost_attrs *loa;
1120         int                      rc;
1121
1122         fld_range_set_any(range);
1123         rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(fid), range);
1124         if (rc == 0) {
1125                 if (fld_range_is_ost(range))
1126                         return 1;
1127
1128                 return 0;
1129         }
1130
1131         loa = &lfsck_env_info(env)->lti_loa;
1132         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, loa, sizeof(*loa)),
1133                           XATTR_NAME_LMA);
1134         if (rc >= (int)sizeof(struct lustre_mdt_attrs)) {
1135                 lustre_lma_swab(&loa->loa_lma);
1136
1137                 return loa->loa_lma.lma_compat & LMAC_FID_ON_OST ? 1 : 0;
1138         }
1139
1140         rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_FID);
1141
1142         return rc > 0;
1143 }
1144
1145 static struct lfsck_layout_seq *
1146 lfsck_layout_seq_lookup(struct lfsck_layout_slave_data *llsd, __u64 seq)
1147 {
1148         struct lfsck_layout_seq *lls;
1149
1150         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
1151                 if (lls->lls_seq == seq)
1152                         return lls;
1153
1154                 if (lls->lls_seq > seq)
1155                         return NULL;
1156         }
1157
1158         return NULL;
1159 }
1160
1161 static void
1162 lfsck_layout_seq_insert(struct lfsck_layout_slave_data *llsd,
1163                         struct lfsck_layout_seq *lls)
1164 {
1165         struct lfsck_layout_seq *tmp;
1166         struct list_head        *pos = &llsd->llsd_seq_list;
1167
1168         list_for_each_entry(tmp, &llsd->llsd_seq_list, lls_list) {
1169                 if (lls->lls_seq < tmp->lls_seq) {
1170                         pos = &tmp->lls_list;
1171                         break;
1172                 }
1173         }
1174         list_add_tail(&lls->lls_list, pos);
1175 }
1176
1177 static int
1178 lfsck_layout_lastid_create(const struct lu_env *env,
1179                            struct lfsck_instance *lfsck,
1180                            struct dt_object *obj)
1181 {
1182         struct lfsck_thread_info *info   = lfsck_env_info(env);
1183         struct lu_attr           *la     = &info->lti_la;
1184         struct dt_object_format  *dof    = &info->lti_dof;
1185         struct lfsck_bookmark    *bk     = &lfsck->li_bookmark_ram;
1186         struct dt_device         *dt     = lfsck_obj2dev(obj);
1187         struct thandle           *th;
1188         __u64                     lastid = 0;
1189         loff_t                    pos    = 0;
1190         int                       rc;
1191         ENTRY;
1192
1193         if (bk->lb_param & LPF_DRYRUN)
1194                 return 0;
1195
1196         memset(la, 0, sizeof(*la));
1197         la->la_mode = S_IFREG |  S_IRUGO | S_IWUSR;
1198         la->la_valid = LA_MODE | LA_UID | LA_GID;
1199         memset(dof, 0, sizeof(*dof));
1200         dof->dof_type = dt_mode_to_dft(S_IFREG);
1201
1202         th = dt_trans_create(env, dt);
1203         if (IS_ERR(th))
1204                 GOTO(log, rc = PTR_ERR(th));
1205
1206         rc = dt_declare_create(env, obj, la, NULL, dof, th);
1207         if (rc != 0)
1208                 GOTO(stop, rc);
1209
1210         rc = dt_declare_record_write(env, obj,
1211                                      lfsck_buf_get(env, &lastid,
1212                                                    sizeof(lastid)),
1213                                      pos, th);
1214         if (rc != 0)
1215                 GOTO(stop, rc);
1216
1217         rc = dt_trans_start_local(env, dt, th);
1218         if (rc != 0)
1219                 GOTO(stop, rc);
1220
1221         dt_write_lock(env, obj, 0);
1222         if (likely(dt_object_exists(obj) == 0)) {
1223                 rc = dt_create(env, obj, la, NULL, dof, th);
1224                 if (rc == 0)
1225                         rc = dt_record_write(env, obj,
1226                                 lfsck_buf_get(env, &lastid, sizeof(lastid)),
1227                                 &pos, th);
1228         }
1229         dt_write_unlock(env, obj);
1230
1231         GOTO(stop, rc);
1232
1233 stop:
1234         dt_trans_stop(env, dt, th);
1235
1236 log:
1237         CDEBUG(D_LFSCK, "%s: layout LFSCK will create LAST_ID for <seq> "
1238                "%#llx: rc = %d\n",
1239                lfsck_lfsck2name(lfsck), fid_seq(lfsck_dto2fid(obj)), rc);
1240
1241         return rc;
1242 }
1243
1244 static int
1245 lfsck_layout_lastid_reload(const struct lu_env *env,
1246                            struct lfsck_component *com,
1247                            struct lfsck_layout_seq *lls)
1248 {
1249         __u64   lastid;
1250         loff_t  pos     = 0;
1251         int     rc;
1252
1253         dt_read_lock(env, lls->lls_lastid_obj, 0);
1254         rc = dt_record_read(env, lls->lls_lastid_obj,
1255                             lfsck_buf_get(env, &lastid, sizeof(lastid)), &pos);
1256         dt_read_unlock(env, lls->lls_lastid_obj);
1257         if (unlikely(rc != 0))
1258                 return rc;
1259
1260         lastid = le64_to_cpu(lastid);
1261         if (lastid < lls->lls_lastid_known) {
1262                 struct lfsck_instance   *lfsck  = com->lc_lfsck;
1263                 struct lfsck_layout     *lo     = com->lc_file_ram;
1264
1265                 lls->lls_lastid = lls->lls_lastid_known;
1266                 lls->lls_dirty = 1;
1267                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1268                         LASSERT(lfsck->li_out_notify != NULL);
1269
1270                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1271                                              LE_LASTID_REBUILDING);
1272                         lo->ll_flags |= LF_CRASHED_LASTID;
1273
1274                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds crashed "
1275                                "LAST_ID file (1) for the sequence %#llx"
1276                                ", old value %llu, known value %llu\n",
1277                                lfsck_lfsck2name(lfsck), lls->lls_seq,
1278                                lastid, lls->lls_lastid);
1279                 }
1280         } else if (lastid >= lls->lls_lastid) {
1281                 lls->lls_lastid = lastid;
1282                 lls->lls_dirty = 0;
1283         }
1284
1285         return 0;
1286 }
1287
1288 static int
1289 lfsck_layout_lastid_store(const struct lu_env *env,
1290                           struct lfsck_component *com)
1291 {
1292         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1293         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
1294         struct dt_device                *dt     = lfsck->li_bottom;
1295         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
1296         struct lfsck_layout_seq         *lls;
1297         struct thandle                  *th;
1298         __u64                            lastid;
1299         int                              rc     = 0;
1300         int                              rc1    = 0;
1301
1302         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
1303                 loff_t pos = 0;
1304
1305                 if (!lls->lls_dirty)
1306                         continue;
1307
1308                 CDEBUG(D_LFSCK, "%s: layout LFSCK will sync the LAST_ID for "
1309                        "<seq> %#llx as <oid> %llu\n",
1310                        lfsck_lfsck2name(lfsck), lls->lls_seq, lls->lls_lastid);
1311
1312                 if (bk->lb_param & LPF_DRYRUN) {
1313                         lls->lls_dirty = 0;
1314                         continue;
1315                 }
1316
1317                 th = dt_trans_create(env, dt);
1318                 if (IS_ERR(th)) {
1319                         rc1 = PTR_ERR(th);
1320                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to store "
1321                                "the LAST_ID for <seq> %#llx(1): rc = %d\n",
1322                                lfsck_lfsck2name(com->lc_lfsck),
1323                                lls->lls_seq, rc1);
1324                         continue;
1325                 }
1326
1327                 lastid = cpu_to_le64(lls->lls_lastid);
1328                 rc = dt_declare_record_write(env, lls->lls_lastid_obj,
1329                                              lfsck_buf_get(env, &lastid,
1330                                                            sizeof(lastid)),
1331                                              pos, th);
1332                 if (rc != 0)
1333                         goto stop;
1334
1335                 rc = dt_trans_start_local(env, dt, th);
1336                 if (rc != 0)
1337                         goto stop;
1338
1339                 dt_write_lock(env, lls->lls_lastid_obj, 0);
1340                 rc = dt_record_write(env, lls->lls_lastid_obj,
1341                                      lfsck_buf_get(env, &lastid,
1342                                      sizeof(lastid)), &pos, th);
1343                 dt_write_unlock(env, lls->lls_lastid_obj);
1344                 if (rc == 0)
1345                         lls->lls_dirty = 0;
1346
1347 stop:
1348                 dt_trans_stop(env, dt, th);
1349                 if (rc != 0) {
1350                         rc1 = rc;
1351                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to store "
1352                                "the LAST_ID for <seq> %#llx(2): rc = %d\n",
1353                                lfsck_lfsck2name(com->lc_lfsck),
1354                                lls->lls_seq, rc1);
1355                 }
1356         }
1357
1358         return rc1;
1359 }
1360
1361 static int
1362 lfsck_layout_lastid_load(const struct lu_env *env,
1363                          struct lfsck_component *com,
1364                          struct lfsck_layout_seq *lls)
1365 {
1366         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1367         struct lfsck_layout     *lo     = com->lc_file_ram;
1368         struct lu_fid           *fid    = &lfsck_env_info(env)->lti_fid;
1369         struct dt_object        *obj;
1370         loff_t                   pos    = 0;
1371         int                      rc;
1372         ENTRY;
1373
1374         lu_last_id_fid(fid, lls->lls_seq, lfsck_dev_idx(lfsck));
1375         obj = dt_locate(env, lfsck->li_bottom, fid);
1376         if (IS_ERR(obj))
1377                 RETURN(PTR_ERR(obj));
1378
1379         /* LAST_ID crashed, to be rebuilt */
1380         if (dt_object_exists(obj) == 0) {
1381                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1382                         LASSERT(lfsck->li_out_notify != NULL);
1383
1384                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1385                                              LE_LASTID_REBUILDING);
1386                         lo->ll_flags |= LF_CRASHED_LASTID;
1387
1388                         CDEBUG(D_LFSCK, "%s: layout LFSCK cannot find the "
1389                                "LAST_ID file for sequence %#llx\n",
1390                                lfsck_lfsck2name(lfsck), lls->lls_seq);
1391
1392                         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY4) &&
1393                             cfs_fail_val > 0) {
1394                                 struct ptlrpc_thread *thread =
1395                                         &lfsck->li_thread;
1396
1397                                 up_write(&com->lc_sem);
1398                                 wait_event_idle_timeout(
1399                                         thread->t_ctl_waitq,
1400                                         !thread_is_running(thread),
1401                                         cfs_time_seconds(cfs_fail_val));
1402                                 down_write(&com->lc_sem);
1403                         }
1404                 }
1405
1406                 rc = lfsck_layout_lastid_create(env, lfsck, obj);
1407         } else {
1408                 dt_read_lock(env, obj, 0);
1409                 rc = dt_read(env, obj,
1410                         lfsck_buf_get(env, &lls->lls_lastid, sizeof(__u64)),
1411                         &pos);
1412                 dt_read_unlock(env, obj);
1413                 if (rc != 0 && rc != sizeof(__u64))
1414                         GOTO(out, rc = (rc > 0 ? -EFAULT : rc));
1415
1416                 if (rc == 0 && !(lo->ll_flags & LF_CRASHED_LASTID)) {
1417                         LASSERT(lfsck->li_out_notify != NULL);
1418
1419                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1420                                              LE_LASTID_REBUILDING);
1421                         lo->ll_flags |= LF_CRASHED_LASTID;
1422
1423                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds invalid "
1424                                "LAST_ID file for the sequence %#llx"
1425                                ": rc = %d\n",
1426                                lfsck_lfsck2name(lfsck), lls->lls_seq, rc);
1427                 }
1428
1429                 lls->lls_lastid = le64_to_cpu(lls->lls_lastid);
1430                 rc = 0;
1431         }
1432
1433         GOTO(out, rc);
1434
1435 out:
1436         if (rc != 0)
1437                 lfsck_object_put(env, obj);
1438         else
1439                 lls->lls_lastid_obj = obj;
1440
1441         return rc;
1442 }
1443
1444 static void lfsck_layout_record_failure(const struct lu_env *env,
1445                                         struct lfsck_instance *lfsck,
1446                                         struct lfsck_layout *lo)
1447 {
1448         __u64 cookie;
1449
1450         lo->ll_objs_failed_phase1++;
1451         cookie = lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
1452                                                         lfsck->li_di_oit);
1453         if (lo->ll_pos_first_inconsistent == 0 ||
1454             lo->ll_pos_first_inconsistent < cookie) {
1455                 lo->ll_pos_first_inconsistent = cookie;
1456
1457                 CDEBUG(D_LFSCK, "%s: layout LFSCK hit first non-repaired "
1458                        "inconsistency at the pos [%llu]\n",
1459                        lfsck_lfsck2name(lfsck),
1460                        lo->ll_pos_first_inconsistent);
1461         }
1462 }
1463
1464 static int lfsck_layout_double_scan_result(const struct lu_env *env,
1465                                            struct lfsck_component *com,
1466                                            int rc)
1467 {
1468         struct lfsck_instance   *lfsck = com->lc_lfsck;
1469         struct lfsck_layout     *lo    = com->lc_file_ram;
1470
1471         CDEBUG(D_LFSCK, "%s: layout LFSCK double scan: rc = %d\n",
1472                lfsck_lfsck2name(lfsck), rc);
1473
1474         down_write(&com->lc_sem);
1475         lo->ll_run_time_phase2 += ktime_get_seconds() -
1476                                   com->lc_time_last_checkpoint;
1477         lo->ll_time_last_checkpoint = ktime_get_real_seconds();
1478         lo->ll_objs_checked_phase2 += com->lc_new_checked;
1479
1480         if (rc > 0) {
1481                 if (lo->ll_flags & LF_INCOMPLETE) {
1482                         lo->ll_status = LS_PARTIAL;
1483                 } else {
1484                         if (lfsck->li_master) {
1485                                 struct lfsck_assistant_data *lad = com->lc_data;
1486
1487                                 if (test_bit(LAD_INCOMPLETE, &lad->lad_flags))
1488                                         lo->ll_status = LS_PARTIAL;
1489                                 else
1490                                         lo->ll_status = LS_COMPLETED;
1491                         } else {
1492                                 lo->ll_status = LS_COMPLETED;
1493                         }
1494                 }
1495                 lo->ll_flags &= ~LF_SCANNED_ONCE;
1496                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN))
1497                         lo->ll_flags &= ~LF_INCONSISTENT;
1498                 lo->ll_time_last_complete = lo->ll_time_last_checkpoint;
1499                 lo->ll_success_count++;
1500         } else if (rc == 0) {
1501                 if (lfsck->li_status != 0)
1502                         lo->ll_status = lfsck->li_status;
1503                 else
1504                         lo->ll_status = LS_STOPPED;
1505         } else {
1506                 lo->ll_status = LS_FAILED;
1507         }
1508
1509         rc = lfsck_layout_store(env, com);
1510         up_write(&com->lc_sem);
1511
1512         CDEBUG(D_LFSCK, "%s: layout LFSCK double scan result %u: rc = %d\n",
1513                lfsck_lfsck2name(lfsck), lo->ll_status, rc);
1514
1515         return rc;
1516 }
1517
1518 static int lfsck_layout_trans_stop(const struct lu_env *env,
1519                                    struct dt_device *dev,
1520                                    struct thandle *handle, int result)
1521 {
1522         int rc;
1523
1524         /* XXX: If there is something worng or it needs to repair nothing,
1525          *      then notify the lower to stop the modification. Currently,
1526          *      we use th_result for such purpose, that may be replaced by
1527          *      some rollback mechanism in the future. */
1528         handle->th_result = result;
1529         rc = dt_trans_stop(env, dev, handle);
1530         if (result != 0)
1531                 return result > 0 ? 0 : result;
1532
1533         return rc == 0 ? 1 : rc;
1534 }
1535
1536 static int lfsck_layout_ins_dangling_rec(const struct lu_env *env,
1537                                          struct lfsck_component *com,
1538                                          const struct lu_fid *pfid,
1539                                          const struct lu_fid *cfid,
1540                                          __u32 comp_id, __u32 ea_off,
1541                                          __u32 ost_idx)
1542 {
1543         struct lfsck_layout_dangling_key *key = &lfsck_env_info(env)->lti_lldk;
1544         struct lu_fid *rec = &lfsck_env_info(env)->lti_fid3;
1545         struct dt_device *dev;
1546         struct dt_object *obj;
1547         struct thandle *th = NULL;
1548         int idx;
1549         int rc = 0;
1550         ENTRY;
1551
1552         idx = lfsck_sub_trace_file_fid2idx(pfid);
1553         obj = com->lc_sub_trace_objs[idx].lsto_obj;
1554         dev = lfsck_obj2dev(obj);
1555
1556         fid_cpu_to_be(&key->lldk_fid, pfid);
1557         key->lldk_comp_id = cpu_to_be32(comp_id);
1558         key->lldk_ea_off = cpu_to_be32(ea_off);
1559
1560         fid_cpu_to_be(rec, cfid);
1561         rec->f_ver = cpu_to_be32(ost_idx);
1562
1563         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1564
1565         th = dt_trans_create(env, dev);
1566         if (IS_ERR(th))
1567                 GOTO(unlock, rc = PTR_ERR(th));
1568
1569         rc = dt_declare_insert(env, obj,
1570                                (const struct dt_rec *)rec,
1571                                (const struct dt_key *)key, th);
1572         if (rc)
1573                 GOTO(unlock, rc);
1574
1575         rc = dt_trans_start_local(env, dev, th);
1576         if (rc)
1577                 GOTO(unlock, rc);
1578
1579         rc = dt_insert(env, obj, (const struct dt_rec *)rec,
1580                        (const struct dt_key *)key, th);
1581
1582         GOTO(unlock, rc);
1583
1584 unlock:
1585         if (th && !IS_ERR(th))
1586                 dt_trans_stop(env, dev, th);
1587
1588         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1589
1590         CDEBUG(D_LFSCK, "%s: insert the paris "DFID" => "DFID", comp_id = %u, "
1591                "ea_off = %u, ost_idx = %u, into the trace file for further "
1592                "dangling check: rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
1593                PFID(pfid), PFID(cfid), comp_id, ea_off, ost_idx, rc);
1594
1595         return rc;
1596 }
1597
1598 static int lfsck_layout_del_dangling_rec(const struct lu_env *env,
1599                                          struct lfsck_component *com,
1600                                          const struct lu_fid *fid,
1601                                          __u32 comp_id, __u32 ea_off)
1602 {
1603         struct lfsck_layout_dangling_key *key = &lfsck_env_info(env)->lti_lldk;
1604         struct dt_device *dev;
1605         struct dt_object *obj;
1606         struct thandle *th = NULL;
1607         int idx;
1608         int rc = 0;
1609         ENTRY;
1610
1611         idx = lfsck_sub_trace_file_fid2idx(fid);
1612         obj = com->lc_sub_trace_objs[idx].lsto_obj;
1613         dev = lfsck_obj2dev(obj);
1614
1615         fid_cpu_to_be(&key->lldk_fid, fid);
1616         key->lldk_comp_id = cpu_to_be32(comp_id);
1617         key->lldk_ea_off = cpu_to_be32(ea_off);
1618
1619         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1620
1621         th = dt_trans_create(env, dev);
1622         if (IS_ERR(th))
1623                 GOTO(unlock, rc = PTR_ERR(th));
1624
1625         rc = dt_declare_delete(env, obj, (const struct dt_key *)key, th);
1626         if (rc)
1627                 GOTO(unlock, rc);
1628
1629         rc = dt_trans_start_local(env, dev, th);
1630         if (rc)
1631                 GOTO(unlock, rc);
1632
1633         rc = dt_delete(env, obj, (const struct dt_key *)key, th);
1634
1635         GOTO(unlock, rc);
1636
1637 unlock:
1638         if (th && !IS_ERR(th))
1639                 dt_trans_stop(env, dev, th);
1640
1641         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1642
1643         CDEBUG(D_LFSCK, "%s: delete the dangling record for "DFID
1644                ", comp_id = %u, ea_off = %u from the trace file: rc = %d\n",
1645                lfsck_lfsck2name(com->lc_lfsck), PFID(fid), comp_id, ea_off, rc);
1646
1647         return rc;
1648 }
1649
1650 /**
1651  * Get the system default stripe size.
1652  *
1653  * \param[in] env       pointer to the thread context
1654  * \param[in] lfsck     pointer to the lfsck instance
1655  * \param[out] size     pointer to the default stripe size
1656  *
1657  * \retval              0 for success
1658  * \retval              negative error number on failure
1659  */
1660 static int lfsck_layout_get_def_stripesize(const struct lu_env *env,
1661                                            struct lfsck_instance *lfsck,
1662                                            __u32 *size)
1663 {
1664         struct lov_user_md      *lum = &lfsck_env_info(env)->lti_lum;
1665         struct dt_object        *root;
1666         int                      rc;
1667
1668         root = dt_locate(env, lfsck->li_next, &lfsck->li_local_root_fid);
1669         if (IS_ERR(root))
1670                 return PTR_ERR(root);
1671
1672         /* Get the default stripe size via xattr_get on the backend root. */
1673         rc = dt_xattr_get(env, root, lfsck_buf_get(env, lum, sizeof(*lum)),
1674                           XATTR_NAME_LOV);
1675         if (rc > 0) {
1676                 /* The lum->lmm_stripe_size is LE mode. The *size also
1677                  * should be LE mode. So it is unnecessary to convert. */
1678                 *size = lum->lmm_stripe_size;
1679                 rc = 0;
1680         } else if (unlikely(rc == 0)) {
1681                 rc = -EINVAL;
1682         }
1683
1684         lfsck_object_put(env, root);
1685
1686         return rc;
1687 }
1688
1689 /**
1690  * \retval       +1: repaired
1691  * \retval        0: did nothing
1692  * \retval      -ve: on error
1693  */
1694 static int lfsck_layout_refill_lovea(const struct lu_env *env,
1695                                      struct lfsck_instance *lfsck,
1696                                      struct thandle *handle,
1697                                      struct dt_object *parent,
1698                                      const struct lu_fid *cfid,
1699                                      struct lu_buf *buf,
1700                                      struct lov_mds_md_v1 *lmm,
1701                                      struct lov_ost_data_v1 *slot,
1702                                      int fl, __u32 ost_idx, int size)
1703 {
1704         struct ost_id           *oi     = &lfsck_env_info(env)->lti_oi;
1705         struct lu_buf            ea_buf;
1706         int                      rc;
1707         __u32                    magic;
1708         __u32                    pattern;
1709         __u16                    count;
1710         ENTRY;
1711
1712         magic = le32_to_cpu(lmm->lmm_magic);
1713         pattern = le32_to_cpu(lmm->lmm_pattern);
1714         count = le16_to_cpu(lmm->lmm_stripe_count);
1715
1716         fid_to_ostid(cfid, oi);
1717         ostid_cpu_to_le(oi, &slot->l_ost_oi);
1718         slot->l_ost_gen = cpu_to_le32(0);
1719         slot->l_ost_idx = cpu_to_le32(ost_idx);
1720
1721         if (pattern & LOV_PATTERN_F_HOLE) {
1722                 struct lov_ost_data_v1 *objs;
1723                 int                     i;
1724
1725                 if (magic == LOV_MAGIC_V1)
1726                         objs = &lmm->lmm_objects[0];
1727                 else
1728                         objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
1729                 for (i = 0; i < count; i++, objs++) {
1730                         if (lovea_slot_is_dummy(objs))
1731                                 break;
1732                 }
1733
1734                 /* If the @slot is the last dummy slot to be refilled,
1735                  * then drop LOV_PATTERN_F_HOLE from lmm::lmm_pattern. */
1736                 if (i == count) {
1737                         lmm->lmm_pattern =
1738                                 cpu_to_le32(pattern & ~LOV_PATTERN_F_HOLE);
1739
1740                         CDEBUG(D_LFSCK, "%s: remove layout HOLE for "DFID
1741                                ": parent "DFID"\n", lfsck_lfsck2name(lfsck),
1742                                PFID(cfid), PFID(lfsck_dto2fid(parent)));
1743                 }
1744         }
1745
1746         lfsck_buf_init(&ea_buf, buf->lb_buf, size);
1747         rc = dt_xattr_set(env, parent, &ea_buf, XATTR_NAME_LOV, fl, handle);
1748         if (rc == 0)
1749                 rc = 1;
1750
1751         RETURN(rc);
1752 }
1753
1754 static struct lov_ost_data_v1 *
1755 __lfsck_layout_new_v1_lovea(struct lov_mds_md_v1 *lmm,
1756                             const struct lu_fid *pfid,
1757                             __u32 stripe_size, __u32 ea_off,
1758                             __u32 pattern, __u16 count)
1759 {
1760         lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
1761         lmm->lmm_pattern = cpu_to_le32(pattern);
1762         fid_to_lmm_oi(pfid, &lmm->lmm_oi);
1763         lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
1764         lmm->lmm_stripe_size = cpu_to_le32(stripe_size);
1765         lmm->lmm_stripe_count = cpu_to_le16(count);
1766         lmm->lmm_layout_gen = cpu_to_le16(1);
1767         memset(&lmm->lmm_objects[0], 0,
1768                sizeof(struct lov_ost_data_v1) * count);
1769
1770         return &lmm->lmm_objects[ea_off];
1771 }
1772
1773 static int lfsck_layout_new_v1_lovea(const struct lu_env *env,
1774                                      struct lfsck_instance *lfsck,
1775                                      struct ost_layout *ol,
1776                                      struct dt_object *parent,
1777                                      struct lu_buf *buf, __u32 ea_off,
1778                                      struct lov_mds_md_v1 **lmm,
1779                                      struct lov_ost_data_v1 **objs)
1780 {
1781         int size;
1782         __u32 stripe_size = ol->ol_stripe_size;
1783         __u32 pattern = LOV_PATTERN_RAID0;
1784         __u16 count;
1785
1786         if (ol->ol_stripe_count != 0)
1787                 count = ol->ol_stripe_count;
1788         else
1789                 count = ea_off + 1;
1790
1791         size = lov_mds_md_size(count, LOV_MAGIC_V1);
1792         LASSERTF(buf->lb_len >= size,
1793                  "buffer len %d is less than real size %d\n",
1794                  (int)buf->lb_len, size);
1795
1796         if (stripe_size == 0) {
1797                 int rc;
1798
1799                 rc = lfsck_layout_get_def_stripesize(env, lfsck, &stripe_size);
1800                 if (rc)
1801                         return rc;
1802         }
1803
1804         *lmm = buf->lb_buf;
1805         if (ol->ol_stripe_count > 1 ||
1806             (ol->ol_stripe_count == 0 && ea_off != 0)) {
1807                 pattern |= LOV_PATTERN_F_HOLE;
1808                 memset(&(*lmm)->lmm_objects[0], 0,
1809                        count * sizeof(struct lov_ost_data_v1));
1810         }
1811
1812         *objs = __lfsck_layout_new_v1_lovea(*lmm, lfsck_dto2fid(parent),
1813                                 stripe_size, ea_off, pattern, count);
1814
1815         return size;
1816 }
1817
1818 static int lfsck_layout_new_comp_lovea(const struct lu_env *env,
1819                                        struct lu_orphan_rec_v3 *rec,
1820                                        struct dt_object *parent,
1821                                        struct lu_buf *buf, __u32 ea_off,
1822                                        struct lov_mds_md_v1 **lmm,
1823                                        struct lov_ost_data_v1 **objs)
1824 {
1825         struct ost_layout *ol = &rec->lor_layout;
1826         struct lov_comp_md_v1 *lcm;
1827         struct lov_comp_md_entry_v1 *lcme;
1828         __u32 pattern = LOV_PATTERN_RAID0;
1829         __u32 offset = sizeof(*lcm) + sizeof(*lcme);
1830         int lcme_size = lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
1831         int size = offset + lcme_size;
1832
1833         LASSERTF(buf->lb_len >= size,
1834                  "buffer len %d is less than real size %d\n",
1835                  (int)buf->lb_len, size);
1836
1837         lcm = buf->lb_buf;
1838         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1839         lcm->lcm_size = cpu_to_le32(size);
1840         if (rec->lor_range) {
1841                 lcm->lcm_layout_gen = cpu_to_le32(rec->lor_layout_version +
1842                                                   rec->lor_range);
1843                 lcm->lcm_flags = cpu_to_le16(LCM_FL_WRITE_PENDING);
1844         } else if (rec->lor_layout_version) {
1845                 lcm->lcm_layout_gen = cpu_to_le32(rec->lor_layout_version +
1846                                                   rec->lor_range);
1847                 lcm->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1848         } else {
1849                 lcm->lcm_layout_gen = cpu_to_le32(1);
1850                 lcm->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1851         }
1852         lcm->lcm_entry_count = cpu_to_le16(1);
1853         /* Currently, we do not know how many mirrors will be, set it as zero
1854          * at the beginning. It will be updated when more mirrors are found. */
1855         lcm->lcm_mirror_count = 0;
1856
1857         lcme = &lcm->lcm_entries[0];
1858         lcme->lcme_id = cpu_to_le32(ol->ol_comp_id);
1859         lcme->lcme_flags = cpu_to_le32(LCME_FL_INIT);
1860         lcme->lcme_extent.e_start = cpu_to_le64(ol->ol_comp_start);
1861         lcme->lcme_extent.e_end = cpu_to_le64(ol->ol_comp_end);
1862         lcme->lcme_offset = cpu_to_le32(offset);
1863         lcme->lcme_size = cpu_to_le32(lcme_size);
1864         lcme->lcme_layout_gen = lcm->lcm_layout_gen;
1865         if (ol->ol_stripe_count > 1)
1866                 pattern |= LOV_PATTERN_F_HOLE;
1867
1868         *lmm = buf->lb_buf + offset;
1869         *objs = __lfsck_layout_new_v1_lovea(*lmm, lfsck_dto2fid(parent),
1870                                             ol->ol_stripe_size, ea_off,
1871                                             pattern, ol->ol_stripe_count);
1872
1873         return size;
1874 }
1875
1876 static void lfsck_layout_update_lcm(struct lov_comp_md_v1 *lcm,
1877                                     struct lov_comp_md_entry_v1 *lcme,
1878                                     __u32 version, __u32 range)
1879 {
1880         struct lov_comp_md_entry_v1 *tmp;
1881         __u64 start = le64_to_cpu(lcme->lcme_extent.e_start);
1882         __u64 end = le64_to_cpu(lcme->lcme_extent.e_end);
1883         __u32 gen = version + range;
1884         __u32 tmp_gen;
1885         int i;
1886         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
1887         __u16 flags = le16_to_cpu(lcm->lcm_flags);
1888
1889         if (!gen)
1890                 gen = 1;
1891         lcme->lcme_layout_gen = cpu_to_le32(gen);
1892         if (le32_to_cpu(lcm->lcm_layout_gen) < gen)
1893                 lcm->lcm_layout_gen = cpu_to_le32(gen);
1894
1895         if (range)
1896                 lcm->lcm_flags = cpu_to_le16(LCM_FL_WRITE_PENDING);
1897         else if (flags == LCM_FL_NONE && le16_to_cpu(lcm->lcm_mirror_count) > 0)
1898                 lcm->lcm_flags = cpu_to_le16(LCM_FL_RDONLY);
1899
1900         for (i = 0; i < count; i++) {
1901                 tmp = &lcm->lcm_entries[i];
1902                 if (le64_to_cpu(tmp->lcme_extent.e_end) <= start)
1903                         continue;
1904
1905                 if (le64_to_cpu(tmp->lcme_extent.e_start) >= end)
1906                         continue;
1907
1908                 if (le32_to_cpu(tmp->lcme_flags) & LCME_FL_STALE)
1909                         continue;
1910
1911                 tmp_gen = le32_to_cpu(tmp->lcme_layout_gen);
1912                 /* "lcme_layout_gen == 0" but without LCME_FL_STALE flag,
1913                  * then it should be the latest version of all mirrors. */
1914                 if (tmp_gen == 0 || tmp_gen > gen) {
1915                         lcme->lcme_flags = cpu_to_le32(
1916                                 le32_to_cpu(lcme->lcme_flags) | LCME_FL_STALE);
1917                         break;
1918                 }
1919
1920                 if (tmp_gen < gen)
1921                         tmp->lcme_flags = cpu_to_le32(
1922                                 le32_to_cpu(tmp->lcme_flags) | LCME_FL_STALE);
1923         }
1924 }
1925
1926 static int lfsck_layout_add_comp(const struct lu_env *env,
1927                                  struct lfsck_instance *lfsck,
1928                                  struct thandle *handle,
1929                                  struct lu_orphan_rec_v3 *rec,
1930                                  struct dt_object *parent,
1931                                  const struct lu_fid *cfid,
1932                                  struct lu_buf *buf, __u32 ost_idx,
1933                                  __u32 ea_off, int pos, bool new_mirror)
1934 {
1935         struct ost_layout *ol = &rec->lor_layout;
1936         struct lov_comp_md_v1 *lcm = buf->lb_buf;
1937         struct lov_comp_md_entry_v1 *lcme;
1938         struct lov_mds_md_v1 *lmm;
1939         struct lov_ost_data_v1 *objs;
1940         int added = sizeof(*lcme) +
1941                     lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
1942         int size = le32_to_cpu(lcm->lcm_size) + added;
1943         int rc;
1944         int i;
1945         __u32 offset;
1946         __u32 pattern = LOV_PATTERN_RAID0;
1947         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
1948         ENTRY;
1949
1950         lu_buf_check_and_grow(buf, size);
1951         /* set the lcm again because lu_buf_check_and_grow() may
1952          * have reallocated the buf. */
1953         lcm = buf->lb_buf;
1954         lcm->lcm_size = cpu_to_le32(size);
1955         lcm->lcm_entry_count = cpu_to_le16(count + 1);
1956         if (new_mirror)
1957                 le16_add_cpu(&lcm->lcm_mirror_count, 1);
1958
1959         /* 1. Move the component bodies from [pos, count-1] to [pos+1, count]
1960          *    with distance of 'added'. */
1961         if (pos < count) {
1962                 size = 0;
1963                 for (i = pos; i < count; i++) {
1964                         lcme = &lcm->lcm_entries[i];
1965                         size += le32_to_cpu(lcme->lcme_size);
1966                 }
1967
1968                 offset = le32_to_cpu(lcm->lcm_entries[pos].lcme_offset);
1969                 memmove(buf->lb_buf + offset + added,
1970                         buf->lb_buf + offset, size);
1971         }
1972
1973         size = 0;
1974         /* 2. Move the component header [0, pos-1] to [0, pos-1] with distance
1975          *    of 'sizeof(struct lov_comp_md_entry_v1)' */
1976         if (pos > 0) {
1977                 for (i = 0; i < pos; i++) {
1978                         lcme = &lcm->lcm_entries[i];
1979                         size += le32_to_cpu(lcme->lcme_size);
1980                 }
1981
1982                 offset = le32_to_cpu(lcm->lcm_entries[0].lcme_offset);
1983                 memmove(buf->lb_buf + offset + sizeof(*lcme),
1984                         buf->lb_buf + offset, size);
1985         }
1986
1987         /* 3. Recalculate the enter offset for the component [pos, count-1] */
1988         for (i = count - 1; i >= pos; i--) {
1989                 lcm->lcm_entries[i + 1] = lcm->lcm_entries[i];
1990                 lcm->lcm_entries[i + 1].lcme_offset =
1991                         cpu_to_le32(le32_to_cpu(lcm->lcm_entries[i + 1].
1992                                                 lcme_offset) + added);
1993         }
1994
1995         /* 4. Recalculate the enter offset for the component [0, pos) */
1996         for (i = 0; i < pos; i++) {
1997                 lcm->lcm_entries[i].lcme_offset =
1998                         cpu_to_le32(le32_to_cpu(lcm->lcm_entries[i].
1999                                                 lcme_offset) + sizeof(*lcme));
2000         }
2001
2002         offset = sizeof(*lcm) + sizeof(*lcme) * (count + 1) + size;
2003         /* 4. Insert the new component header (entry) at the slot 'pos'. */
2004         lcme = &lcm->lcm_entries[pos];
2005         lcme->lcme_id = cpu_to_le32(ol->ol_comp_id);
2006         lcme->lcme_flags = cpu_to_le32(LCME_FL_INIT);
2007         lcme->lcme_extent.e_start = cpu_to_le64(ol->ol_comp_start);
2008         lcme->lcme_extent.e_end = cpu_to_le64(ol->ol_comp_end);
2009         lcme->lcme_offset = cpu_to_le32(offset);
2010         lcme->lcme_size = cpu_to_le32(lov_mds_md_size(ol->ol_stripe_count,
2011                                                       LOV_MAGIC_V1));
2012
2013         if (ol->ol_stripe_count > 1)
2014                 pattern |= LOV_PATTERN_F_HOLE;
2015
2016         lmm = buf->lb_buf + offset;
2017         /* 5. Insert teh new component body at the 'offset'. */
2018         objs = __lfsck_layout_new_v1_lovea(lmm, lfsck_dto2fid(parent),
2019                                            ol->ol_stripe_size, ea_off,
2020                                            pattern, ol->ol_stripe_count);
2021
2022         /* 6. Update mirror related flags and version. */
2023         lfsck_layout_update_lcm(lcm, lcme, rec->lor_layout_version,
2024                                 rec->lor_range);
2025
2026         rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid, buf,
2027                                        lmm, objs, LU_XATTR_REPLACE, ost_idx,
2028                                        le32_to_cpu(lcm->lcm_size));
2029
2030         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant add new COMP for "
2031                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2032                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2033                "comp_end %llu, layout version %u, range %u, "
2034                "%s LOV EA hole: rc = %d\n",
2035                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2036                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2037                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
2038                rec->lor_layout_version, rec->lor_range,
2039                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2040                "with" : "without", rc);
2041
2042         RETURN(rc);
2043 }
2044
2045 static int lfsck_layout_extend_v1v3_lovea(const struct lu_env *env,
2046                                           struct lfsck_instance *lfsck,
2047                                           struct thandle *handle,
2048                                           struct ost_layout *ol,
2049                                           struct dt_object *parent,
2050                                           const struct lu_fid *cfid,
2051                                           struct lu_buf *buf, __u32 ost_idx,
2052                                           __u32 ea_off)
2053 {
2054         struct lov_mds_md_v1 *lmm = buf->lb_buf;
2055         struct lov_ost_data_v1 *objs;
2056         __u16 count = le16_to_cpu(lmm->lmm_stripe_count);
2057         __u32 magic = le32_to_cpu(lmm->lmm_magic);
2058         int size;
2059         int gap;
2060         int rc;
2061         ENTRY;
2062
2063         /* The original LOVEA maybe re-generated via old filter_fid, at
2064          * that time, we do not know the stripe count and stripe size. */
2065         if (ol->ol_stripe_count > count)
2066                 count = ol->ol_stripe_count;
2067         if (ol->ol_stripe_size != 0 &&
2068             ol->ol_stripe_size != le32_to_cpu(lmm->lmm_stripe_size))
2069                 lmm->lmm_stripe_size = cpu_to_le32(ol->ol_stripe_size);
2070
2071         if (magic == LOV_MAGIC_V1)
2072                 objs = &lmm->lmm_objects[count];
2073         else
2074                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[count];
2075
2076         gap = ea_off - count;
2077         if (gap >= 0)
2078                 count = ea_off + 1;
2079
2080         size = lov_mds_md_size(count, magic);
2081         LASSERTF(buf->lb_len >= size,
2082                  "buffer len %d is less than real size %d\n",
2083                  (int)buf->lb_len, size);
2084
2085         if (gap > 0) {
2086                 memset(objs, 0, gap * sizeof(*objs));
2087                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_HOLE);
2088         }
2089
2090         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
2091         lmm->lmm_stripe_count = cpu_to_le16(count);
2092         objs += gap;
2093
2094         rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid, buf,
2095                                 lmm, objs, LU_XATTR_REPLACE, ost_idx, size);
2096
2097         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant extend layout EA for "
2098                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2099                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2100                "comp_end %llu, %s LOV EA hole: rc = %d\n",
2101                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2102                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2103                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
2104                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2105                "with" : "without", rc);
2106
2107         RETURN(rc);
2108 }
2109
2110 /**
2111  * \retval       +1: repaired
2112  * \retval        0: did nothing
2113  * \retval      -ve: on error
2114  */
2115 static int lfsck_layout_update_lovea(const struct lu_env *env,
2116                                      struct lfsck_instance *lfsck,
2117                                      struct thandle *handle,
2118                                      struct lu_orphan_rec_v3 *rec,
2119                                      struct dt_object *parent,
2120                                      const struct lu_fid *cfid,
2121                                      struct lu_buf *buf, int fl,
2122                                      __u32 ost_idx, __u32 ea_off)
2123 {
2124         struct ost_layout *ol = &rec->lor_layout;
2125         struct lov_mds_md_v1 *lmm = NULL;
2126         struct lov_ost_data_v1 *objs = NULL;
2127         int rc = 0;
2128         ENTRY;
2129
2130         if (ol->ol_comp_id != 0)
2131                 rc = lfsck_layout_new_comp_lovea(env, rec, parent, buf, ea_off,
2132                                                  &lmm, &objs);
2133         else
2134                 rc = lfsck_layout_new_v1_lovea(env, lfsck, &rec->lor_layout,
2135                                                parent, buf, ea_off, &lmm,
2136                                                &objs);
2137         if (rc > 0)
2138                 rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid,
2139                                                buf, lmm, objs, fl, ost_idx, rc);
2140
2141         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant created layout EA for "
2142                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2143                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2144                "comp_end %llu, layout version %u, range %u, fl %d, "
2145                "%s LOV EA hole: rc = %d\n",
2146                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2147                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2148                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
2149                rec->lor_layout_version, rec->lor_range, fl,
2150                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2151                "with" : "without", rc);
2152
2153         RETURN(rc);
2154 }
2155
2156 static int __lfsck_layout_update_pfid(const struct lu_env *env,
2157                                       struct dt_object *child,
2158                                       const struct lu_fid *pfid,
2159                                       const struct ost_layout *ol, __u32 offset,
2160                                       __u32 version, __u32 range)
2161 {
2162         struct dt_device        *dev    = lfsck_obj2dev(child);
2163         struct filter_fid       *ff     = &lfsck_env_info(env)->lti_ff;
2164         struct thandle          *handle;
2165         struct lu_buf            buf    = { NULL };
2166         int                      rc;
2167
2168         ff->ff_parent.f_seq = cpu_to_le64(pfid->f_seq);
2169         ff->ff_parent.f_oid = cpu_to_le32(pfid->f_oid);
2170         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
2171          * MDT-object's FID::f_ver, instead it is the OST-object index in its
2172          * parent MDT-object's layout EA. */
2173         ff->ff_parent.f_stripe_idx = cpu_to_le32(offset);
2174         ost_layout_cpu_to_le(&ff->ff_layout, ol);
2175         ff->ff_layout_version = cpu_to_le32(version);
2176         ff->ff_range = cpu_to_le32(range);
2177         lfsck_buf_init(&buf, ff, sizeof(*ff));
2178
2179         handle = dt_trans_create(env, dev);
2180         if (IS_ERR(handle))
2181                 RETURN(PTR_ERR(handle));
2182
2183         rc = dt_declare_xattr_set(env, child, &buf, XATTR_NAME_FID, 0, handle);
2184         if (rc != 0)
2185                 GOTO(stop, rc);
2186
2187         rc = dt_trans_start_local(env, dev, handle);
2188         if (rc != 0)
2189                 GOTO(stop, rc);
2190
2191         rc = dt_xattr_set(env, child, &buf, XATTR_NAME_FID, 0, handle);
2192
2193         GOTO(stop, rc);
2194
2195 stop:
2196         dt_trans_stop(env, dev, handle);
2197
2198         return rc;
2199 }
2200
2201 /**
2202  * \retval       +1: repaired
2203  * \retval        0: did nothing
2204  * \retval      -ve: on error
2205  */
2206 static int lfsck_layout_update_pfid(const struct lu_env *env,
2207                                     struct lfsck_component *com,
2208                                     struct dt_object *parent,
2209                                     struct lu_fid *cfid,
2210                                     struct dt_device *cdev,
2211                                     struct lu_orphan_rec_v3 *rec, __u32 ea_off)
2212 {
2213         struct dt_object        *child;
2214         int                      rc     = 0;
2215         ENTRY;
2216
2217         child = lfsck_object_find_by_dev(env, cdev, cfid);
2218         if (IS_ERR(child))
2219                 RETURN(PTR_ERR(child));
2220
2221         rc = __lfsck_layout_update_pfid(env, child,
2222                                         lu_object_fid(&parent->do_lu),
2223                                         &rec->lor_layout, ea_off,
2224                                         rec->lor_layout_version,
2225                                         rec->lor_range);
2226         lfsck_object_put(env, child);
2227
2228         RETURN(rc == 0 ? 1 : rc);
2229 }
2230
2231 static int lfsck_lovea_size(struct ost_layout *ol, __u32 ea_off)
2232 {
2233         if (ol->ol_comp_id != 0)
2234                 return sizeof(struct lov_comp_md_v1) +
2235                        sizeof(struct lov_comp_md_entry_v1) +
2236                        lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
2237
2238         if (ol->ol_stripe_count != 0)
2239                 return lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
2240
2241         return lov_mds_md_size(ea_off + 1, LOV_MAGIC_V1);
2242 }
2243
2244 /**
2245  * This function will create the MDT-object with the given (partial) LOV EA.
2246  *
2247  * Under some data corruption cases, the MDT-object of the file may be lost,
2248  * but its OST-objects, or some of them are there. The layout LFSCK needs to
2249  * re-create the MDT-object with the orphan OST-object(s) information.
2250  *
2251  * On the other hand, the LFSCK may has created some OST-object for repairing
2252  * dangling LOV EA reference, but as the LFSCK processing, it may find that
2253  * the old OST-object is there and should replace the former new created OST
2254  * object. Unfortunately, some others have modified such newly created object.
2255  * To keep the data (both new and old), the LFSCK will create MDT-object with
2256  * new FID to reference the original OST-object.
2257  *
2258  * \param[in] env       pointer to the thread context
2259  * \param[in] com       pointer to the lfsck component
2260  * \param[in] ltd       pointer to target device descriptor
2261  * \param[in] rec       pointer to the record for the orphan OST-object
2262  * \param[in] cfid      pointer to FID for the orphan OST-object
2263  * \param[in] infix     additional information, such as the FID for original
2264  *                      MDT-object and the stripe offset in the LOV EA
2265  * \param[in] type      the type for describing why the orphan MDT-object is
2266  *                      created. The rules are as following:
2267  *
2268  *  type "C":           Multiple OST-objects claim the same MDT-object and the
2269  *                      same slot in the layout EA. Then the LFSCK will create
2270  *                      new MDT-object(s) to hold the conflict OST-object(s).
2271  *
2272  *  type "N":           The orphan OST-object does not know which one was the
2273  *                      real parent MDT-object, so the LFSCK uses new FID for
2274  *                      its parent MDT-object.
2275  *
2276  *  type "R":           The orphan OST-object knows its parent MDT-object FID,
2277  *                      but does not know the position (the file name) in the
2278  *                      layout.
2279  *
2280  *  type "D":           The MDT-object is a directory, it may knows its parent
2281  *                      but because there is no valid linkEA, the LFSCK cannot
2282  *                      know where to put it back to the namespace.
2283  *  type "O":           The MDT-object has no linkEA, and there is no name
2284  *                      entry that references the MDT-object.
2285  *
2286  *  type "P":           The orphan object to be created was a parent directory
2287  *                      of some MDT-object which linkEA shows that the @orphan
2288  *                      object is missing.
2289  *
2290  * The orphan name will be like:
2291  * ${FID}-${infix}-${type}-${conflict_version}
2292  *
2293  * \param[in] ea_off    the stripe offset in the LOV EA
2294  *
2295  * \retval              positive on repaired something
2296  * \retval              0 if needs to repair nothing
2297  * \retval              negative error number on failure
2298  */
2299 static int lfsck_layout_recreate_parent(const struct lu_env *env,
2300                                         struct lfsck_component *com,
2301                                         struct lfsck_tgt_desc *ltd,
2302                                         struct lu_orphan_rec_v3 *rec,
2303                                         struct lu_fid *cfid,
2304                                         const char *infix,
2305                                         const char *type,
2306                                         __u32 ea_off)
2307 {
2308         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2309         struct dt_insert_rec            *dtrec  = &info->lti_dt_rec;
2310         char                            *name   = info->lti_key;
2311         struct lu_attr                  *la     = &info->lti_la2;
2312         struct dt_object_format         *dof    = &info->lti_dof;
2313         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2314         struct lu_fid                   *pfid   = &rec->lor_rec.lor_fid;
2315         struct lu_fid                   *tfid   = &info->lti_fid3;
2316         struct dt_device                *dev    = lfsck->li_bottom;
2317         struct dt_object                *lpf    = lfsck->li_lpf_obj;
2318         struct dt_object                *pobj   = NULL;
2319         struct dt_object                *cobj   = NULL;
2320         struct thandle                  *th     = NULL;
2321         struct lu_buf                   *ea_buf = &info->lti_big_buf;
2322         struct lu_buf                    lov_buf;
2323         struct lfsck_lock_handle        *llh    = &info->lti_llh;
2324         struct linkea_data               ldata  = { NULL };
2325         struct lu_buf                    linkea_buf;
2326         const struct lu_name            *pname;
2327         int                              size   = 0;
2328         int                              idx    = 0;
2329         int                              rc     = 0;
2330         ENTRY;
2331
2332         if (unlikely(lpf == NULL))
2333                 GOTO(log, rc = -ENXIO);
2334
2335         /* We use two separated transactions to repair the inconsistency.
2336          *
2337          * 1) create the MDT-object locally.
2338          * 2) update the OST-object's PFID EA if necessary.
2339          *
2340          * If 1) succeed, but 2) failed, then the OST-object's PFID EA will be
2341          * updated when the layout LFSCK run next time.
2342          *
2343          * If 1) failed, but 2) succeed, then such MDT-object will be re-created
2344          * when the layout LFSCK run next time. */
2345
2346         if (fid_is_zero(pfid)) {
2347                 rc = lfsck_fid_alloc(env, lfsck, pfid, false);
2348                 if (rc != 0)
2349                         GOTO(log, rc);
2350
2351                 cobj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, cfid);
2352                 if (IS_ERR(cobj))
2353                         GOTO(log, rc = PTR_ERR(cobj));
2354         }
2355
2356         pobj = lfsck_object_find_by_dev(env, dev, pfid);
2357         if (IS_ERR(pobj))
2358                 GOTO(log, rc = PTR_ERR(pobj));
2359
2360         LASSERT(infix != NULL);
2361         LASSERT(type != NULL);
2362
2363         memset(la, 0, sizeof(*la));
2364         la->la_uid = rec->lor_rec.lor_uid;
2365         la->la_gid = rec->lor_rec.lor_gid;
2366         la->la_mode = S_IFREG | S_IRUSR;
2367         la->la_valid = LA_MODE | LA_UID | LA_GID;
2368
2369         memset(dof, 0, sizeof(*dof));
2370         dof->dof_type = dt_mode_to_dft(S_IFREG);
2371         /* Because the dof->dof_reg.striped = 0, the LOD will not create
2372          * the stripe(s). The LFSCK will specify the LOV EA via
2373          * lfsck_layout_update_lovea(). */
2374
2375         size = lfsck_lovea_size(&rec->lor_layout, ea_off);
2376         if (ea_buf->lb_len < size) {
2377                 lu_buf_realloc(ea_buf, size);
2378                 if (ea_buf->lb_buf == NULL)
2379                         GOTO(log, rc = -ENOMEM);
2380         }
2381
2382 again:
2383         do {
2384                 snprintf(name, NAME_MAX, DFID"%s-%s-%d", PFID(pfid), infix,
2385                          type, idx++);
2386                 rc = dt_lookup(env, lfsck->li_lpf_obj, (struct dt_rec *)tfid,
2387                                (const struct dt_key *)name);
2388                 if (rc != 0 && rc != -ENOENT)
2389                         GOTO(log, rc);
2390         } while (rc == 0);
2391
2392         rc = lfsck_lock(env, lfsck, lfsck->li_lpf_obj, name, llh,
2393                         MDS_INODELOCK_UPDATE, LCK_PW);
2394         if (rc != 0)
2395                 GOTO(log, rc);
2396
2397         /* Re-check whether the name conflict with othrs after taken
2398          * the ldlm lock. */
2399         rc = dt_lookup(env, lfsck->li_lpf_obj, (struct dt_rec *)tfid,
2400                        (const struct dt_key *)name);
2401         if (unlikely(rc == 0)) {
2402                 lfsck_unlock(llh);
2403                 goto again;
2404         }
2405
2406         if (rc != -ENOENT)
2407                 GOTO(unlock, rc);
2408
2409         pname = lfsck_name_get_const(env, name, strlen(name));
2410         rc = linkea_links_new(&ldata, &lfsck_env_info(env)->lti_linkea_buf,
2411                               pname, lfsck_dto2fid(lfsck->li_lpf_obj));
2412         if (rc != 0)
2413                 GOTO(unlock, rc);
2414
2415         /* The 1st transaction. */
2416         th = dt_trans_create(env, dev);
2417         if (IS_ERR(th))
2418                 GOTO(unlock, rc = PTR_ERR(th));
2419
2420         rc = dt_declare_create(env, pobj, la, NULL, dof, th);
2421         if (rc != 0)
2422                 GOTO(stop, rc);
2423
2424         lfsck_buf_init(&lov_buf, ea_buf->lb_buf, size);
2425         rc = dt_declare_xattr_set(env, pobj, &lov_buf, XATTR_NAME_LOV,
2426                                   LU_XATTR_CREATE, th);
2427         if (rc != 0)
2428                 GOTO(stop, rc);
2429
2430         dtrec->rec_fid = pfid;
2431         dtrec->rec_type = S_IFREG;
2432         rc = dt_declare_insert(env, lpf,
2433                                (const struct dt_rec *)dtrec,
2434                                (const struct dt_key *)name, th);
2435         if (rc != 0)
2436                 GOTO(stop, rc);
2437
2438         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
2439                        ldata.ld_leh->leh_len);
2440         rc = dt_declare_xattr_set(env, pobj, &linkea_buf,
2441                                   XATTR_NAME_LINK, 0, th);
2442         if (rc != 0)
2443                 GOTO(stop, rc);
2444
2445         rc = dt_trans_start_local(env, dev, th);
2446         if (rc != 0)
2447                 GOTO(stop, rc);
2448
2449         dt_write_lock(env, pobj, 0);
2450         rc = dt_create(env, pobj, la, NULL, dof, th);
2451         if (rc == 0)
2452                 rc = lfsck_layout_update_lovea(env, lfsck, th, rec, pobj, cfid,
2453                         &lov_buf, LU_XATTR_CREATE, ltd->ltd_index, ea_off);
2454         dt_write_unlock(env, pobj);
2455         if (rc < 0)
2456                 GOTO(stop, rc);
2457
2458         rc = dt_insert(env, lpf, (const struct dt_rec *)dtrec,
2459                        (const struct dt_key *)name, th);
2460         if (rc != 0)
2461                 GOTO(stop, rc);
2462
2463         rc = dt_xattr_set(env, pobj, &linkea_buf, XATTR_NAME_LINK, 0, th);
2464         if (rc == 0 && cobj != NULL) {
2465                 dt_trans_stop(env, dev, th);
2466                 th = NULL;
2467
2468                 /* The 2nd transaction. */
2469                 rc = __lfsck_layout_update_pfid(env, cobj, pfid,
2470                                                 &rec->lor_layout, ea_off,
2471                                                 rec->lor_layout_version,
2472                                                 rec->lor_range);
2473         }
2474
2475         GOTO(stop, rc);
2476
2477 stop:
2478         if (th != NULL)
2479                 dt_trans_stop(env, dev, th);
2480
2481 unlock:
2482         lfsck_unlock(llh);
2483
2484 log:
2485         if (cobj != NULL && !IS_ERR(cobj))
2486                 lfsck_object_put(env, cobj);
2487         if (pobj != NULL && !IS_ERR(pobj))
2488                 lfsck_object_put(env, pobj);
2489
2490         if (rc < 0)
2491                 CDEBUG(D_LFSCK, "%s layout LFSCK assistant failed to "
2492                        "recreate the lost MDT-object: parent "DFID
2493                        ", child "DFID", OST-index %u, stripe-index %u, "
2494                        "infix %s, type %s: rc = %d\n",
2495                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
2496                        ltd->ltd_index, ea_off, infix, type, rc);
2497
2498         return rc >= 0 ? 1 : rc;
2499 }
2500
2501 static int lfsck_layout_master_conditional_destroy(const struct lu_env *env,
2502                                                    struct lfsck_component *com,
2503                                                    const struct lu_fid *fid,
2504                                                    __u32 index)
2505 {
2506         struct lfsck_thread_info *info  = lfsck_env_info(env);
2507         struct lfsck_request     *lr    = &info->lti_lr;
2508         struct lfsck_instance    *lfsck = com->lc_lfsck;
2509         struct lfsck_tgt_desc    *ltd;
2510         struct ptlrpc_request    *req;
2511         struct lfsck_request     *tmp;
2512         struct obd_export        *exp;
2513         int                       rc    = 0;
2514         ENTRY;
2515
2516         ltd = lfsck_tgt_get(&lfsck->li_ost_descs, index);
2517         if (unlikely(ltd == NULL))
2518                 RETURN(-ENXIO);
2519
2520         exp = ltd->ltd_exp;
2521         if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
2522                 GOTO(put, rc = -EOPNOTSUPP);
2523
2524         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
2525         if (req == NULL)
2526                 GOTO(put, rc = -ENOMEM);
2527
2528         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
2529         if (rc != 0) {
2530                 ptlrpc_request_free(req);
2531
2532                 GOTO(put, rc);
2533         }
2534
2535         memset(lr, 0, sizeof(*lr));
2536         lr->lr_event = LE_CONDITIONAL_DESTROY;
2537         lr->lr_active = LFSCK_TYPE_LAYOUT;
2538         lr->lr_fid = *fid;
2539
2540         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
2541         *tmp = *lr;
2542         ptlrpc_request_set_replen(req);
2543
2544         rc = ptlrpc_queue_wait(req);
2545         ptlrpc_req_finished(req);
2546
2547         GOTO(put, rc);
2548
2549 put:
2550         lfsck_tgt_put(ltd);
2551
2552         return rc;
2553 }
2554
2555 static int lfsck_layout_slave_conditional_destroy(const struct lu_env *env,
2556                                                   struct lfsck_component *com,
2557                                                   struct lfsck_request *lr)
2558 {
2559         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2560         struct lu_attr                  *la     = &info->lti_la;
2561         union ldlm_policy_data          *policy = &info->lti_policy;
2562         struct ldlm_res_id              *resid  = &info->lti_resid;
2563         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2564         struct dt_device                *dev    = lfsck->li_bottom;
2565         struct lu_fid                   *fid    = &lr->lr_fid;
2566         struct dt_object                *obj;
2567         struct thandle                  *th     = NULL;
2568         struct lustre_handle             lh     = { 0 };
2569         __u64                            flags  = 0;
2570         int                              rc     = 0;
2571         ENTRY;
2572
2573         obj = lfsck_object_find_by_dev(env, dev, fid);
2574         if (IS_ERR(obj))
2575                 RETURN(PTR_ERR(obj));
2576
2577         dt_read_lock(env, obj, 0);
2578         if (dt_object_exists(obj) == 0 ||
2579             lfsck_is_dead_obj(obj)) {
2580                 dt_read_unlock(env, obj);
2581
2582                 GOTO(put, rc = -ENOENT);
2583         }
2584
2585         /* Get obj's attr without lock firstly. */
2586         rc = dt_attr_get(env, obj, la);
2587         dt_read_unlock(env, obj);
2588         if (rc != 0)
2589                 GOTO(put, rc);
2590
2591         if (likely(la->la_ctime != 0 || la->la_mode & S_ISUID))
2592                 GOTO(put, rc = -ETXTBSY);
2593
2594         /* Acquire extent lock on [0, EOF] to sync with all possible written. */
2595         LASSERT(lfsck->li_namespace != NULL);
2596
2597         memset(policy, 0, sizeof(*policy));
2598         policy->l_extent.end = OBD_OBJECT_EOF;
2599         ost_fid_build_resid(fid, resid);
2600         rc = ldlm_cli_enqueue_local(env, lfsck->li_namespace, resid,
2601                                     LDLM_EXTENT, policy, LCK_EX, &flags,
2602                                     ldlm_blocking_ast, ldlm_completion_ast,
2603                                     NULL, NULL, 0, LVB_T_NONE, NULL, &lh);
2604         if (rc != ELDLM_OK)
2605                 GOTO(put, rc = -EIO);
2606
2607         dt_write_lock(env, obj, 0);
2608         /* Get obj's attr within lock again. */
2609         rc = dt_attr_get(env, obj, la);
2610         if (rc != 0)
2611                 GOTO(unlock, rc);
2612
2613         if (la->la_ctime != 0)
2614                 GOTO(unlock, rc = -ETXTBSY);
2615
2616         th = dt_trans_create(env, dev);
2617         if (IS_ERR(th))
2618                 GOTO(unlock, rc = PTR_ERR(th));
2619
2620         rc = dt_declare_ref_del(env, obj, th);
2621         if (rc != 0)
2622                 GOTO(stop, rc);
2623
2624         rc = dt_declare_destroy(env, obj, th);
2625         if (rc != 0)
2626                 GOTO(stop, rc);
2627
2628         rc = dt_trans_start_local(env, dev, th);
2629         if (rc != 0)
2630                 GOTO(stop, rc);
2631
2632         rc = dt_ref_del(env, obj, th);
2633         if (rc != 0)
2634                 GOTO(stop, rc);
2635
2636         rc = dt_destroy(env, obj, th);
2637         if (rc == 0)
2638                 CDEBUG(D_LFSCK, "%s: layout LFSCK destroyed the empty "
2639                        "OST-object "DFID" that was created for reparing "
2640                        "dangling referenced case. But the original missing "
2641                        "OST-object is found now.\n",
2642                        lfsck_lfsck2name(lfsck), PFID(fid));
2643
2644         GOTO(stop, rc);
2645
2646 stop:
2647         dt_trans_stop(env, dev, th);
2648
2649 unlock:
2650         dt_write_unlock(env, obj);
2651         ldlm_lock_decref(&lh, LCK_EX);
2652
2653 put:
2654         lfsck_object_put(env, obj);
2655
2656         return rc;
2657 }
2658
2659 /**
2660  * Some OST-object has occupied the specified layout EA slot.
2661  * Such OST-object may be generated by the LFSCK when repair
2662  * dangling referenced MDT-object, which can be indicated by
2663  * attr::la_ctime == 0 but without S_ISUID in la_mode. If it
2664  * is true and such OST-object has not been modified yet, we
2665  * will replace it with the orphan OST-object; otherwise the
2666  * LFSCK will create new MDT-object to reference the orphan.
2667  *
2668  * \retval       +1: repaired
2669  * \retval        0: did nothing
2670  * \retval      -ve: on error
2671  */
2672 static int lfsck_layout_conflict_create(const struct lu_env *env,
2673                                         struct lfsck_component *com,
2674                                         struct lfsck_tgt_desc *ltd,
2675                                         struct lu_orphan_rec_v3 *rec,
2676                                         struct dt_object *parent,
2677                                         struct lu_fid *cfid,
2678                                         struct lu_buf *ea_buf,
2679                                         struct lov_mds_md_v1 *lmm,
2680                                         struct lov_ost_data_v1 *slot,
2681                                         __u32 ea_off, int lovea_size)
2682 {
2683         struct lfsck_thread_info *info          = lfsck_env_info(env);
2684         struct lu_fid            *cfid2         = &info->lti_fid2;
2685         struct ost_id            *oi            = &info->lti_oi;
2686         struct dt_device         *dev           = lfsck_obj2dev(parent);
2687         struct thandle           *th            = NULL;
2688         struct lustre_handle      lh            = { 0 };
2689         __u32                     ost_idx2      = le32_to_cpu(slot->l_ost_idx);
2690         int                       rc            = 0;
2691         ENTRY;
2692
2693         while (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val)) {
2694                 if (unlikely(!thread_is_running(&com->lc_lfsck->li_thread)))
2695                         RETURN(0);
2696         }
2697
2698         ostid_le_to_cpu(&slot->l_ost_oi, oi);
2699         rc = ostid_to_fid(cfid2, oi, ost_idx2);
2700         if (rc != 0)
2701                 GOTO(out, rc);
2702
2703         rc = lfsck_ibits_lock(env, com->lc_lfsck, parent, &lh,
2704                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
2705                               LCK_EX);
2706         if (rc != 0)
2707                 GOTO(out, rc);
2708
2709         rc = lfsck_layout_master_conditional_destroy(env, com, cfid2, ost_idx2);
2710
2711         /* If the conflict OST-obejct is not created for fixing dangling
2712          * referenced MDT-object in former LFSCK check/repair, or it has
2713          * been modified by others, then we cannot destroy it. Re-create
2714          * a new MDT-object for the orphan OST-object. */
2715         if (rc == -ETXTBSY) {
2716                 /* No need the layout lock on the original parent. */
2717                 lfsck_ibits_unlock(&lh, LCK_EX);
2718
2719                 fid_zero(&rec->lor_rec.lor_fid);
2720                 snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2721                          "-"DFID"-%x", PFID(lu_object_fid(&parent->do_lu)),
2722                          ea_off);
2723                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
2724                                                 info->lti_tmpbuf, "C", ea_off);
2725
2726                 RETURN(rc);
2727         }
2728
2729         if (rc != 0 && rc != -ENOENT)
2730                 GOTO(unlock, rc);
2731
2732         th = dt_trans_create(env, dev);
2733         if (IS_ERR(th))
2734                 GOTO(unlock, rc = PTR_ERR(th));
2735
2736         rc = dt_declare_xattr_set(env, parent, ea_buf, XATTR_NAME_LOV,
2737                                   LU_XATTR_REPLACE, th);
2738         if (rc != 0)
2739                 GOTO(stop, rc);
2740
2741         rc = dt_trans_start_local(env, dev, th);
2742         if (rc != 0)
2743                 GOTO(stop, rc);
2744
2745         dt_write_lock(env, parent, 0);
2746         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
2747         rc = lfsck_layout_refill_lovea(env, com->lc_lfsck, th, parent, cfid,
2748                                        ea_buf, lmm, slot, LU_XATTR_REPLACE,
2749                                        ltd->ltd_index, lovea_size);
2750         dt_write_unlock(env, parent);
2751
2752         GOTO(stop, rc);
2753
2754 stop:
2755         dt_trans_stop(env, dev, th);
2756
2757 unlock:
2758         lfsck_ibits_unlock(&lh, LCK_EX);
2759
2760 out:
2761         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant replaced the conflict "
2762                "OST-object "DFID" on the OST %x with the orphan "DFID" on "
2763                "the OST %x: parent "DFID", stripe-index %u: rc = %d\n",
2764                lfsck_lfsck2name(com->lc_lfsck), PFID(cfid2), ost_idx2,
2765                PFID(cfid), ltd->ltd_index, PFID(lfsck_dto2fid(parent)),
2766                ea_off, rc);
2767
2768         return rc >= 0 ? 1 : rc;
2769 }
2770
2771 /**
2772  * \retval       +1: repaired
2773  * \retval        0: did nothing
2774  * \retval      -ve: on error
2775  */
2776 static int lfsck_layout_recreate_lovea(const struct lu_env *env,
2777                                        struct lfsck_component *com,
2778                                        struct lfsck_tgt_desc *ltd,
2779                                        struct lu_orphan_rec_v3 *rec,
2780                                        struct dt_object *parent,
2781                                        struct lu_fid *cfid,
2782                                        __u32 ost_idx, __u32 ea_off)
2783 {
2784         struct lfsck_thread_info *info          = lfsck_env_info(env);
2785         struct lu_buf            *buf           = &info->lti_big_buf;
2786         struct lu_fid            *fid           = &info->lti_fid2;
2787         struct ost_id            *oi            = &info->lti_oi;
2788         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2789         struct dt_device         *dt            = lfsck_obj2dev(parent);
2790         struct lfsck_bookmark    *bk            = &lfsck->li_bookmark_ram;
2791         struct ost_layout        *ol            = &rec->lor_layout;
2792         struct lov_comp_md_v1    *lcm           = NULL;
2793         struct lov_comp_md_entry_v1 *lcme       = NULL;
2794         struct thandle           *handle        = NULL;
2795         size_t                    lovea_size;
2796         struct lov_mds_md_v1     *lmm;
2797         struct lov_ost_data_v1   *objs;
2798         struct lustre_handle      lh            = { 0 };
2799         __u32                     magic;
2800         __u32 flags = 0;
2801         int                       fl            = 0;
2802         int                       rc            = 0;
2803         int                       rc1;
2804         int                       i;
2805         int pos = 0;
2806         __u16 count;
2807         bool locked = false;
2808         bool new_mirror = true;
2809         ENTRY;
2810
2811         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
2812                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
2813                               LCK_EX);
2814         if (rc != 0) {
2815                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant failed to recreate "
2816                        "LOV EA for "DFID": parent "DFID", OST-index %u, "
2817                        "stripe-index %u, comp_id %u, comp_start %llu, "
2818                        "comp_end %llu, layout version %u, range %u: rc = %d\n",
2819                        lfsck_lfsck2name(lfsck), PFID(cfid),
2820                        PFID(lfsck_dto2fid(parent)), ost_idx, ea_off,
2821                        ol->ol_comp_id, ol->ol_comp_start,
2822                        ol->ol_comp_end, rec->lor_layout_version,
2823                        rec->lor_range, rc);
2824
2825                 RETURN(rc);
2826         }
2827
2828 again:
2829         if (locked) {
2830                 dt_write_unlock(env, parent);
2831                 locked = false;
2832         }
2833
2834         if (handle != NULL) {
2835                 dt_trans_stop(env, dt, handle);
2836                 handle = NULL;
2837         }
2838
2839         if (rc < 0)
2840                 GOTO(unlock_layout, rc);
2841
2842         lovea_size = rc;
2843         if (buf->lb_len < lovea_size) {
2844                 lu_buf_realloc(buf, lovea_size);
2845                 if (buf->lb_buf == NULL)
2846                         GOTO(unlock_layout, rc = -ENOMEM);
2847         }
2848
2849         if (!(bk->lb_param & LPF_DRYRUN)) {
2850                 handle = dt_trans_create(env, dt);
2851                 if (IS_ERR(handle))
2852                         GOTO(unlock_layout, rc = PTR_ERR(handle));
2853
2854                 rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
2855                                           fl, handle);
2856                 if (rc != 0)
2857                         GOTO(stop, rc);
2858
2859                 rc = dt_trans_start_local(env, dt, handle);
2860                 if (rc != 0)
2861                         GOTO(stop, rc);
2862         }
2863
2864         dt_write_lock(env, parent, 0);
2865         locked = true;
2866         rc = dt_xattr_get(env, parent, buf, XATTR_NAME_LOV);
2867         if (rc == -ERANGE) {
2868                 rc = dt_xattr_get(env, parent, &LU_BUF_NULL, XATTR_NAME_LOV);
2869                 LASSERT(rc != 0);
2870                 goto again;
2871         } else if (rc == -ENODATA || rc == 0) {
2872                 lovea_size = lfsck_lovea_size(ol, ea_off);
2873                 /* If the declared is not big enough, re-try. */
2874                 if (buf->lb_len < lovea_size) {
2875                         rc = lovea_size;
2876                         goto again;
2877                 }
2878                 fl = LU_XATTR_CREATE;
2879         } else if (rc < 0) {
2880                 GOTO(unlock_parent, rc);
2881         } else if (unlikely(buf->lb_len == 0)) {
2882                 goto again;
2883         } else {
2884                 fl = LU_XATTR_REPLACE;
2885                 lovea_size = rc;
2886         }
2887
2888         if (fl == LU_XATTR_CREATE) {
2889                 if (bk->lb_param & LPF_DRYRUN)
2890                         GOTO(unlock_parent, rc = 1);
2891
2892                 LASSERT(buf->lb_len >= lovea_size);
2893
2894                 rc = lfsck_layout_update_lovea(env, lfsck, handle, rec, parent,
2895                                                cfid, buf, fl, ost_idx, ea_off);
2896
2897                 GOTO(unlock_parent, rc);
2898         }
2899
2900         lmm = buf->lb_buf;
2901         rc1 = lfsck_layout_verify_header(parent, lmm, lovea_size);
2902
2903         /* If the LOV EA crashed, the rebuild it. */
2904         if (rc1 == -EINVAL) {
2905                 if (bk->lb_param & LPF_DRYRUN)
2906                         GOTO(unlock_parent, rc = 1);
2907
2908                 LASSERT(buf->lb_len >= lovea_size);
2909
2910                 rc = lfsck_layout_update_lovea(env, lfsck, handle, rec, parent,
2911                                                cfid, buf, fl, ost_idx, ea_off);
2912
2913                 GOTO(unlock_parent, rc);
2914         }
2915
2916         /* For other unknown magic/pattern, keep the current LOV EA. */
2917         if (rc1 == -EOPNOTSUPP)
2918                 GOTO(unlock_parent, rc1 = 0);
2919
2920         if (rc1)
2921                 GOTO(unlock_parent, rc = rc1);
2922
2923         magic = le32_to_cpu(lmm->lmm_magic);
2924         if (magic == LOV_MAGIC_COMP_V1) {
2925                 __u64 start;
2926                 __u64 end;
2927                 __u16 mirror_id0 = mirror_id_of(ol->ol_comp_id);
2928                 __u16 mirror_id1;
2929
2930                 if (bk->lb_param & LPF_DRYRUN)
2931                         GOTO(unlock_parent, rc = 1);
2932
2933                 lcm = buf->lb_buf;
2934                 count = le16_to_cpu(lcm->lcm_entry_count);
2935                 for (i = 0; i < count; pos = ++i) {
2936                         lcme = &lcm->lcm_entries[i];
2937                         start = le64_to_cpu(lcme->lcme_extent.e_start);
2938                         end = le64_to_cpu(lcme->lcme_extent.e_end);
2939                         mirror_id1 = mirror_id_of(le32_to_cpu(lcme->lcme_id));
2940
2941                         if (mirror_id0 > mirror_id1)
2942                                 continue;
2943
2944                         if (mirror_id0 < mirror_id1)
2945                                 break;
2946
2947                         new_mirror = false;
2948                         if (end <= ol->ol_comp_start)
2949                                 continue;
2950
2951                         if (start >= ol->ol_comp_end)
2952                                 break;
2953
2954                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
2955                         magic = le32_to_cpu(lmm->lmm_magic);
2956                         flags = le32_to_cpu(lcme->lcme_flags);
2957                         goto further;
2958                 }
2959
2960                 rc = lfsck_layout_add_comp(env, lfsck, handle, rec, parent,
2961                                 cfid, buf, ost_idx, ea_off, pos, new_mirror);
2962
2963                 GOTO(unlock_parent, rc);
2964         }
2965
2966 further:
2967         count = le16_to_cpu(lmm->lmm_stripe_count);
2968         if (count == 0)
2969                 GOTO(unlock_parent, rc = -EINVAL);
2970         LASSERT(count > 0);
2971
2972         /* Exceed the current end of MDT-object layout EA. Then extend it. */
2973         if (count <= ea_off) {
2974                 if (bk->lb_param & LPF_DRYRUN)
2975                         GOTO(unlock_parent, rc = 1);
2976
2977                 lovea_size = lov_mds_md_size(ea_off + 1, magic);
2978                 /* If the declared is not big enough, re-try. */
2979                 if (buf->lb_len < lovea_size) {
2980                         rc = lovea_size;
2981                         goto again;
2982                 }
2983
2984                 if (lcm) {
2985                         LASSERT(lcme);
2986
2987                         lcme->lcme_flags = cpu_to_le32(flags | LCME_FL_INIT);
2988                         lfsck_layout_update_lcm(lcm, lcme,
2989                                                 rec->lor_layout_version,
2990                                                 rec->lor_range);
2991                 }
2992
2993                 rc = lfsck_layout_extend_v1v3_lovea(env, lfsck, handle, ol,
2994                                         parent, cfid, buf, ost_idx, ea_off);
2995
2996                 GOTO(unlock_parent, rc);
2997         }
2998
2999         LASSERTF(rc > 0, "invalid rc = %d\n", rc);
3000
3001         if (magic == LOV_MAGIC_V1) {
3002                 objs = &lmm->lmm_objects[0];
3003         } else {
3004                 LASSERT(magic == LOV_MAGIC_V3);
3005                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
3006         }
3007
3008         for (i = 0; i < count; i++, objs++) {
3009                 /* The MDT-object was created via lfsck_layout_recover_create()
3010                  * by others before, and we fill the dummy layout EA. */
3011                 if ((lcme && !(flags & LCME_FL_INIT)) ||
3012                      lovea_slot_is_dummy(objs)) {
3013                         if (i != ea_off)
3014                                 continue;
3015
3016                         if (bk->lb_param & LPF_DRYRUN)
3017                                 GOTO(unlock_parent, rc = 1);
3018
3019                         lmm->lmm_layout_gen =
3020                             cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
3021                         if (lcme) {
3022                                 LASSERT(lcm);
3023
3024                                 if (le32_to_cpu(lmm->lmm_stripe_size) !=
3025                                         ol->ol_stripe_size ||
3026                                     le16_to_cpu(lmm->lmm_stripe_count) !=
3027                                         ol->ol_stripe_count ||
3028                                     le64_to_cpu(lcme->lcme_extent.e_start) !=
3029                                         ol->ol_comp_start ||
3030                                     le64_to_cpu(lcme->lcme_extent.e_end) !=
3031                                         ol->ol_comp_end) {
3032                                         CDEBUG(D_LFSCK, "%s: found invalid "
3033                                         "component for "DFID ": parent "DFID
3034                                         ", stripe-index %u, stripe_size %u, "
3035                                         "stripe_count %u, comp_id %u, "
3036                                         "comp_start %llu, comp_end %llu, "
3037                                         "cur_stripe_size %u, "
3038                                         "cur_stripe_count %u, "
3039                                         "cur_comp_start %llu, "
3040                                         "cur_comp_end %llu\n",
3041                                         lfsck_lfsck2name(lfsck), PFID(cfid),
3042                                         PFID(lfsck_dto2fid(parent)), ea_off,
3043                                         ol->ol_stripe_size,
3044                                         ol->ol_stripe_count, ol->ol_comp_id,
3045                                         ol->ol_comp_start, ol->ol_comp_end,
3046                                         le32_to_cpu(lmm->lmm_stripe_size),
3047                                         le16_to_cpu(lmm->lmm_stripe_count),
3048                                         le64_to_cpu(lcme->lcme_extent.e_start),
3049                                         le64_to_cpu(lcme->lcme_extent.e_end));
3050
3051                                         GOTO(unlock_parent, rc = -EINVAL);
3052                                 }
3053
3054                                 lovea_size = le32_to_cpu(lcm->lcm_size);
3055                                 lcme->lcme_flags = cpu_to_le32(flags |
3056                                                                LCME_FL_INIT);
3057                                 lfsck_layout_update_lcm(lcm, lcme,
3058                                                         rec->lor_layout_version,
3059                                                         rec->lor_range);
3060                         }
3061
3062                         LASSERTF(buf->lb_len >= lovea_size,
3063                                  "buffer len %d is less than real size %d\n",
3064                                  (int)buf->lb_len, (int)lovea_size);
3065
3066                         rc = lfsck_layout_refill_lovea(env, lfsck, handle,
3067                                                 parent, cfid, buf, lmm, objs,
3068                                                 fl, ost_idx, lovea_size);
3069
3070                         CDEBUG(D_LFSCK, "%s layout LFSCK assistant fill "
3071                                "dummy layout slot for "DFID": parent "DFID
3072                                ", OST-index %u, stripe-index %u: rc = %d\n",
3073                                lfsck_lfsck2name(lfsck), PFID(cfid),
3074                                PFID(lfsck_dto2fid(parent)), ost_idx, i, rc);
3075
3076                         GOTO(unlock_parent, rc);
3077                 }
3078
3079                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
3080                 rc = ostid_to_fid(fid, oi, le32_to_cpu(objs->l_ost_idx));
3081                 if (rc != 0) {
3082                         CDEBUG(D_LFSCK, "%s: the parent "DFID" contains "
3083                                "invalid layout EA at the slot %d, index %u\n",
3084                                lfsck_lfsck2name(lfsck),
3085                                PFID(lfsck_dto2fid(parent)), i,
3086                                le32_to_cpu(objs->l_ost_idx));
3087
3088                         GOTO(unlock_parent, rc);
3089                 }
3090
3091                 /* It should be rare case, the slot is there, but the LFSCK
3092                  * does not handle it during the first-phase cycle scanning. */
3093                 if (unlikely(lu_fid_eq(fid, cfid))) {
3094                         if (i == ea_off) {
3095                                 GOTO(unlock_parent, rc = 0);
3096                         } else {
3097                                 /* Rare case that the OST-object index
3098                                  * does not match the parent MDT-object
3099                                  * layout EA. We trust the later one. */
3100                                 if (bk->lb_param & LPF_DRYRUN)
3101                                         GOTO(unlock_parent, rc = 1);
3102
3103                                 dt_write_unlock(env, parent);
3104                                 if (handle != NULL)
3105                                         dt_trans_stop(env, dt, handle);
3106                                 lfsck_ibits_unlock(&lh, LCK_EX);
3107                                 rc = lfsck_layout_update_pfid(env, com, parent,
3108                                                         cfid, ltd->ltd_tgt,
3109                                                         rec, i);
3110
3111                                 CDEBUG(D_LFSCK, "%s layout LFSCK assistant "
3112                                        "updated OST-object's pfid for "DFID
3113                                        ": parent "DFID", OST-index %u, "
3114                                        "stripe-index %u: rc = %d\n",
3115                                        lfsck_lfsck2name(lfsck), PFID(cfid),
3116                                        PFID(lfsck_dto2fid(parent)),
3117                                        ltd->ltd_index, i, rc);
3118
3119                                 RETURN(rc);
3120                         }
3121                 }
3122         }
3123
3124         /* The MDT-object exists, but related layout EA slot is occupied
3125          * by others. */
3126         if (bk->lb_param & LPF_DRYRUN)
3127                 GOTO(unlock_parent, rc = 1);
3128
3129         dt_write_unlock(env, parent);
3130         if (handle != NULL)
3131                 dt_trans_stop(env, dt, handle);
3132         lfsck_ibits_unlock(&lh, LCK_EX);
3133         if (magic == LOV_MAGIC_V1)
3134                 objs = &lmm->lmm_objects[ea_off];
3135         else
3136                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[ea_off];
3137         rc = lfsck_layout_conflict_create(env, com, ltd, rec, parent, cfid,
3138                                           buf, lmm, objs, ea_off, lovea_size);
3139
3140         RETURN(rc);
3141
3142 unlock_parent:
3143         if (locked)
3144                 dt_write_unlock(env, parent);
3145
3146 stop:
3147         if (handle != NULL)
3148                 dt_trans_stop(env, dt, handle);
3149
3150 unlock_layout:
3151         lfsck_ibits_unlock(&lh, LCK_EX);
3152
3153         return rc;
3154 }
3155
3156 static int lfsck_layout_scan_orphan_one(const struct lu_env *env,
3157                                         struct lfsck_component *com,
3158                                         struct lfsck_tgt_desc *ltd,
3159                                         struct lu_orphan_rec_v3 *rec,
3160                                         struct lu_fid *cfid)
3161 {
3162         struct lfsck_layout     *lo     = com->lc_file_ram;
3163         struct lu_fid           *pfid   = &rec->lor_rec.lor_fid;
3164         struct dt_object        *parent = NULL;
3165         __u32                    ea_off = pfid->f_stripe_idx;
3166         int                      rc     = 0;
3167         ENTRY;
3168
3169         if (!fid_is_sane(cfid))
3170                 GOTO(out, rc = -EINVAL);
3171
3172         pfid->f_ver = 0;
3173         if (fid_is_zero(pfid)) {
3174                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
3175                                                   "", "N", ea_off);
3176                 GOTO(out, rc);
3177         }
3178
3179         if (!fid_is_sane(pfid))
3180                 GOTO(out, rc = -EINVAL);
3181
3182         parent = lfsck_object_find_by_dev(env, com->lc_lfsck->li_bottom, pfid);
3183         if (IS_ERR(parent))
3184                 GOTO(out, rc = PTR_ERR(parent));
3185
3186         if (unlikely(dt_object_remote(parent) != 0))
3187                 GOTO(put, rc = -EXDEV);
3188
3189         if (dt_object_exists(parent) == 0) {
3190                 lfsck_object_put(env, parent);
3191                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
3192                                                   "", "R", ea_off);
3193                 GOTO(out, rc);
3194         }
3195
3196         if (!S_ISREG(lu_object_attr(&parent->do_lu)))
3197                 GOTO(put, rc = -EISDIR);
3198
3199         /* The orphan OST-object claims to be the parent's stripe, then
3200          * related dangling record in the trace file is meaningless. */
3201         rc = lfsck_layout_del_dangling_rec(env, com, pfid,
3202                                            rec->lor_layout.ol_comp_id, ea_off);
3203         if (rc && rc != -ENOENT)
3204                 GOTO(put, rc);
3205
3206         rc = lfsck_layout_recreate_lovea(env, com, ltd, rec, parent, cfid,
3207                                          ltd->ltd_index, ea_off);
3208
3209         GOTO(put, rc);
3210
3211 put:
3212         if (rc <= 0)
3213                 lfsck_object_put(env, parent);
3214         else
3215                 /* The layout EA is changed, need to be reloaded next time. */
3216                 dt_object_put_nocache(env, parent);
3217
3218 out:
3219         down_write(&com->lc_sem);
3220         com->lc_new_scanned++;
3221         com->lc_new_checked++;
3222         if (rc > 0) {
3223                 lo->ll_objs_repaired[LLIT_ORPHAN - 1]++;
3224                 rc = 0;
3225         } else if (rc < 0) {
3226                 lo->ll_objs_failed_phase2++;
3227         }
3228         up_write(&com->lc_sem);
3229
3230         return rc;
3231 }
3232
3233 static int lfsck_layout_scan_orphan(const struct lu_env *env,
3234                                     struct lfsck_component *com,
3235                                     struct lfsck_tgt_desc *ltd)
3236 {
3237         struct lfsck_assistant_data     *lad    = com->lc_data;
3238         struct lfsck_instance           *lfsck  = com->lc_lfsck;
3239         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
3240         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3241         struct lu_fid                   *fid    = &info->lti_fid;
3242         struct dt_object                *obj;
3243         const struct dt_it_ops          *iops;
3244         struct dt_it                    *di;
3245         int                              rc     = 0;
3246         ENTRY;
3247
3248         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant starts the orphan "
3249                "scanning for OST%04x\n",
3250                lfsck_lfsck2name(lfsck), ltd->ltd_index);
3251
3252         if (cfs_bitmap_check(lad->lad_bitmap, ltd->ltd_index)) {
3253                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant skip the orphan "
3254                        "scanning for OST%04x\n",
3255                        lfsck_lfsck2name(lfsck), ltd->ltd_index);
3256
3257                 RETURN(0);
3258         }
3259
3260         fid->f_seq = fid_idif_seq(0, ltd->ltd_index);
3261         fid->f_oid = fid->f_ver = 0;
3262
3263         obj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, fid);
3264         if (unlikely(IS_ERR(obj)))
3265                 GOTO(log, rc = PTR_ERR(obj));
3266
3267         rc = obj->do_ops->do_index_try(env, obj,
3268                                        &dt_lfsck_layout_orphan_features);
3269         if (rc != 0)
3270                 GOTO(put, rc);
3271
3272         iops = &obj->do_index_ops->dio_it;
3273         di = iops->init(env, obj, 0);
3274         if (IS_ERR(di))
3275                 GOTO(put, rc = PTR_ERR(di));
3276
3277         rc = iops->load(env, di, 0);
3278         if (rc == -ESRCH) {
3279                 /* -ESRCH means that the orphan OST-objects rbtree has been
3280                  * cleanup because of the OSS server restart or other errors. */
3281                 lfsck_lad_set_bitmap(env, com, ltd->ltd_index);
3282                 GOTO(fini, rc);
3283         }
3284
3285         if (rc == 0)
3286                 rc = iops->next(env, di);
3287         else if (rc > 0)
3288                 rc = 0;
3289
3290         if (rc < 0)
3291                 GOTO(fini, rc);
3292
3293         if (rc > 0)
3294                 GOTO(fini, rc = 0);
3295
3296         do {
3297                 struct dt_key           *key;
3298                 struct lu_orphan_rec_v3 *rec = &info->lti_rec;
3299
3300                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
3301                     unlikely(!thread_is_running(&lfsck->li_thread)))
3302                         break;
3303
3304                 key = iops->key(env, di);
3305                 com->lc_fid_latest_scanned_phase2 = *(struct lu_fid *)key;
3306                 /* Remote target OST may be runnning old LFSCK */
3307                 memset(rec, 0, sizeof(*rec));
3308                 rc = iops->rec(env, di, (struct dt_rec *)rec, 0);
3309                 if (rc == 0)
3310                         rc = lfsck_layout_scan_orphan_one(env, com, ltd, rec,
3311                                         &com->lc_fid_latest_scanned_phase2);
3312                 if (rc != 0 && bk->lb_param & LPF_FAILOUT)
3313                         GOTO(fini, rc);
3314
3315                 lfsck_control_speed_by_self(com);
3316                 do {
3317                         rc = iops->next(env, di);
3318                 } while (rc < 0 && !(bk->lb_param & LPF_FAILOUT));
3319         } while (rc == 0);
3320
3321         GOTO(fini, rc);
3322
3323 fini:
3324         iops->put(env, di);
3325         iops->fini(env, di);
3326 put:
3327         lfsck_object_put(env, obj);
3328
3329 log:
3330         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant finished the orphan "
3331                "scanning for OST%04x: rc = %d\n",
3332                lfsck_lfsck2name(lfsck), ltd->ltd_index, rc);
3333
3334         return rc > 0 ? 0 : rc;
3335 }
3336
3337 static int lfsck_lov2layout(struct lov_mds_md_v1 *lmm, struct filter_fid *ff,
3338                             __u32 comp_id)
3339 {
3340         struct ost_layout *ol = &ff->ff_layout;
3341         __u32 magic = le32_to_cpu(lmm->lmm_magic);
3342         int rc = 0;
3343         ENTRY;
3344
3345         if (magic == LOV_MAGIC_V1 || magic == LOV_MAGIC_V3) {
3346                 ol->ol_stripe_size = lmm->lmm_stripe_size;
3347                 ol->ol_stripe_count = lmm->lmm_stripe_count;
3348                 ol->ol_comp_start = 0;
3349                 ol->ol_comp_end = 0;
3350                 ol->ol_comp_id = 0;
3351                 ff->ff_layout_version = 0;
3352                 ff->ff_range = 0;
3353         } else if (magic == LOV_MAGIC_COMP_V1) {
3354                 struct lov_comp_md_v1 *lcm = (struct lov_comp_md_v1 *)lmm;
3355                 struct lov_comp_md_entry_v1 *lcme = NULL;
3356                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3357                 int i;
3358
3359                 for (i = 0; i < count; i++) {
3360                         lcme = &lcm->lcm_entries[i];
3361                         if (le32_to_cpu(lcme->lcme_id) == comp_id) {
3362                                 LASSERT(le32_to_cpu(lcme->lcme_flags) &
3363                                         LCME_FL_INIT);
3364
3365                                 break;
3366                         }
3367                 }
3368
3369                 /* The comp has been removed, do nothing. */
3370                 if (i == count)
3371                         GOTO(out, rc = 1);
3372
3373                 lmm = (void *)lmm + le32_to_cpu(lcme->lcme_offset);
3374                 ol->ol_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
3375                 ol->ol_stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
3376                 ol->ol_comp_start = le64_to_cpu(lcme->lcme_extent.e_start);
3377                 ol->ol_comp_end = le64_to_cpu(lcme->lcme_extent.e_end);
3378                 ol->ol_comp_id = le32_to_cpu(lcme->lcme_id);
3379                 ff->ff_layout_version = le32_to_cpu(lcme->lcme_layout_gen);
3380                 ff->ff_range = 0;
3381         } else {
3382                 GOTO(out, rc = -EINVAL);
3383         }
3384
3385         EXIT;
3386
3387 out:
3388         return rc;
3389 }
3390
3391 /**
3392  * Repair the MDT-object with dangling LOV EA reference.
3393  *
3394  * we need to repair the inconsistency according to the users' requirement:
3395  *
3396  * 1) Keep the inconsistency there and report the inconsistency case,
3397  *    then give the chance to the application to find related issues,
3398  *    and the users can make the decision about how to handle it with
3399  *    more human knownledge. (by default)
3400  *
3401  * 2) Re-create the missing OST-object with the FID/owner information.
3402  *
3403  * \param[in] env       pointer to the thread context
3404  * \param[in] com       the layout LFSCK component
3405  * \param[in] parent    the MDT-object with dangling LOV EA reference
3406  * \param[in] child     the OST-object to be created
3407  * \param[in] comp_id   the component ID of the OST-object in the LOV EA
3408  * \param[in] ea_off    the offset of the OST-object in the LOV EA
3409  * \param[in] ost_idx   the index of OST on which the OST-object resides
3410  *
3411  * \retval              +1 for repair successfully
3412  * \retval              0 for did nothing
3413  * \retval              negative error number on failure
3414  */
3415 static int __lfsck_layout_repair_dangling(const struct lu_env *env,
3416                                           struct lfsck_component *com,
3417                                           struct dt_object *parent,
3418                                           struct dt_object *child,
3419                                           __u32 comp_id, __u32 ea_off,
3420                                           __u32 ost_idx, bool log)
3421 {
3422         struct lfsck_thread_info *info = lfsck_env_info(env);
3423         struct filter_fid *ff = &info->lti_ff;
3424         struct dt_object_format *dof = &info->lti_dof;
3425         struct lu_attr *la = &info->lti_la;
3426         struct lfsck_instance *lfsck = com->lc_lfsck;
3427         struct dt_device *dev = lfsck_obj2dev(child);
3428         const struct lu_fid *pfid = lfsck_dto2fid(parent);
3429         const struct lu_fid *cfid = lfsck_dto2fid(child);
3430         struct lu_buf *tbuf = &info->lti_big_buf;
3431         struct thandle *handle;
3432         struct lu_buf *buf;
3433         struct lustre_handle lh = { 0 };
3434         int rc;
3435         ENTRY;
3436
3437         if (!(lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ))
3438                 GOTO(log, rc = 1);
3439
3440         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
3441                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3442                               LCK_EX);
3443         if (rc != 0)
3444                 GOTO(log, rc);
3445
3446         rc = dt_attr_get(env, parent, la);
3447         if (rc != 0)
3448                 GOTO(unlock1, rc);
3449
3450         la->la_mode = S_IFREG | 0666;
3451         la->la_atime = la->la_mtime = la->la_ctime = 0;
3452         la->la_valid = LA_TYPE | LA_MODE | LA_UID | LA_GID |
3453                        LA_ATIME | LA_MTIME | LA_CTIME;
3454         memset(dof, 0, sizeof(*dof));
3455         ff->ff_parent.f_seq = cpu_to_le64(pfid->f_seq);
3456         ff->ff_parent.f_oid = cpu_to_le32(pfid->f_oid);
3457         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
3458          * MDT-object's FID::f_ver, instead it is the OST-object index in its
3459          * parent MDT-object's layout EA. */
3460         ff->ff_parent.f_stripe_idx = cpu_to_le32(ea_off);
3461
3462         rc = lfsck_layout_get_lovea(env, parent, tbuf);
3463         if (unlikely(rc == -ENODATA))
3464                 rc = 0;
3465         if (rc <= 0)
3466                 GOTO(unlock1, rc);
3467
3468         rc = lfsck_lov2layout(tbuf->lb_buf, ff, comp_id);
3469         if (rc)
3470                 GOTO(unlock1, rc);
3471
3472         buf = lfsck_buf_get(env, ff, sizeof(struct filter_fid));
3473         handle = dt_trans_create(env, dev);
3474         if (IS_ERR(handle))
3475                 GOTO(unlock1, rc = PTR_ERR(handle));
3476
3477         rc = dt_declare_create(env, child, la, NULL, dof, handle);
3478         if (rc != 0)
3479                 GOTO(stop, rc);
3480
3481         rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID,
3482                                   LU_XATTR_CREATE, handle);
3483         if (rc != 0)
3484                 GOTO(stop, rc);
3485
3486         rc = dt_trans_start_local(env, dev, handle);
3487         if (rc != 0)
3488                 GOTO(stop, rc);
3489
3490         dt_read_lock(env, parent, 0);
3491         if (unlikely(lfsck_is_dead_obj(parent)))
3492                 GOTO(unlock2, rc = 0);
3493
3494         if (lfsck->li_bookmark_ram.lb_param & LPF_DELAY_CREATE_OSTOBJ) {
3495                 struct ost_id *oi = &info->lti_oi;
3496                 struct lu_fid *tfid = &info->lti_fid2;
3497                 struct lu_buf *lovea = &info->lti_big_buf;
3498                 struct lov_mds_md_v1 *lmm;
3499                 struct lov_ost_data_v1 *objs;
3500                 __u32 magic;
3501                 int count;
3502                 int idx2;
3503
3504                 rc = lfsck_layout_get_lovea(env, parent, lovea);
3505                 if (unlikely(rc == -ENODATA))
3506                         rc = 0;
3507                 if (rc <= 0)
3508                         GOTO(unlock2, rc);
3509
3510                 lmm = lovea->lb_buf;
3511                 magic = le32_to_cpu(lmm->lmm_magic);
3512                 if (magic == LOV_MAGIC_COMP_V1) {
3513                         struct lov_comp_md_v1 *lcm = buf->lb_buf;
3514                         struct lov_comp_md_entry_v1 *lcme;
3515                         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3516                         int i;
3517
3518                         for (i = 0; i < count; i++) {
3519                                 lcme = &lcm->lcm_entries[i];
3520                                 if (le32_to_cpu(lcme->lcme_id) == comp_id) {
3521                                         LASSERT(le32_to_cpu(lcme->lcme_flags) &
3522                                                 LCME_FL_INIT);
3523
3524                                         lmm = lovea->lb_buf +
3525                                                 le32_to_cpu(lcme->lcme_offset);
3526                                         magic = le32_to_cpu(lmm->lmm_magic);
3527                                         goto check;
3528                                 }
3529                         }
3530
3531                         /* Someone removed the component, do nothing. */
3532                         GOTO(unlock2, rc = 0);
3533                 }
3534
3535 check:
3536                 count = le16_to_cpu(lmm->lmm_stripe_count);
3537                 /* Someone changed the LOV EA, do nothing. */
3538                 if (count <= ea_off)
3539                         GOTO(unlock2, rc = 0);
3540
3541                 if (magic == LOV_MAGIC_V1) {
3542                         objs = &lmm->lmm_objects[ea_off];
3543                 } else {
3544                         LASSERT(magic == LOV_MAGIC_V3);
3545
3546                         objs = &((struct lov_mds_md_v3 *)lmm)->\
3547                                                         lmm_objects[ea_off];
3548                 }
3549
3550                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
3551                 idx2 = le32_to_cpu(objs->l_ost_idx);
3552                 rc = ostid_to_fid(tfid, oi, idx2);
3553                 /* Someone changed the LOV EA, do nothing. */
3554                 if (rc != 0 || !lu_fid_eq(tfid, cfid))
3555                         GOTO(unlock2, rc);
3556         }
3557
3558         rc = dt_create(env, child, la, NULL, dof, handle);
3559         if (rc != 0)
3560                 GOTO(unlock2, rc);
3561
3562         rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, LU_XATTR_CREATE,
3563                           handle);
3564
3565         GOTO(unlock2, rc);
3566
3567 unlock2:
3568         dt_read_unlock(env, parent);
3569
3570 stop:
3571         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3572
3573 unlock1:
3574         lfsck_ibits_unlock(&lh, LCK_EX);
3575
3576 log:
3577         if (rc && log)
3578                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant found "
3579                        "dangling reference for: parent "DFID", child "
3580                        DFID", comp_id %u, ea_off %u, ost_idx %u, %s: "
3581                        "rc = %d\n",
3582                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
3583                        comp_id, ea_off, ost_idx,
3584                        (lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ) ?
3585                                 "Create the lost OST-object as required" :
3586                                 "Keep the MDT-object there by default", rc);
3587
3588         return rc;
3589 }
3590
3591 /**
3592  * Repair the MDT-object with dangling LOV EA reference.
3593  *
3594  * Prepare parameters and call __lfsck_layout_repair_dangling()
3595  * to repair the dangling LOV EA reference.
3596  *
3597  * \param[in] env       pointer to the thread context
3598  * \param[in] com       the layout LFSCK component
3599  * \param[in] pfid      the MDT-object's FID
3600  * \param[in] cfid      the FID for the OST-object to be created
3601  * \param[in] comp_id   the component ID of the OST-object in the LOV EA
3602  * \param[in] ea_off    the offset of the OST-object in the LOV EA
3603  * \param[in] ost_idx   the index of OST on which the OST-object resides
3604  *
3605  * \retval              +1 for repair successfully
3606  * \retval              0 for did nothing
3607  * \retval              negative error number on failure
3608  */
3609 static int lfsck_layout_repair_dangling(const struct lu_env *env,
3610                                         struct lfsck_component *com,
3611                                         const struct lu_fid *pfid,
3612                                         const struct lu_fid *cfid,
3613                                         __u32 comp_id, __u32 ea_off,
3614                                         __u32 ost_idx)
3615 {
3616         struct lfsck_instance *lfsck = com->lc_lfsck;
3617         struct dt_object *parent = NULL;
3618         struct dt_object *child = NULL;
3619         struct lfsck_tgt_desc *ltd;
3620         int rc;
3621         ENTRY;
3622
3623         parent = lfsck_object_find_bottom(env, lfsck, pfid);
3624         if (IS_ERR(parent))
3625                 GOTO(log, rc = PTR_ERR(parent));
3626
3627         /* The MDT-object has been removed. */
3628         if (dt_object_exists(parent) == 0)
3629                 GOTO(log, rc = 0);
3630
3631         ltd = lfsck_ltd2tgt(&lfsck->li_ost_descs, ost_idx);
3632         if (unlikely(ltd == NULL))
3633                 GOTO(log, rc = -ENODEV);
3634
3635         child = lfsck_object_find_by_dev(env, ltd->ltd_tgt, cfid);
3636         if (IS_ERR(child))
3637                 GOTO(log, rc = PTR_ERR(child));
3638
3639         /* The OST-object has been created. */
3640         if (unlikely(dt_object_exists(child) != 0))
3641                 GOTO(log, rc = 0);
3642
3643         rc = __lfsck_layout_repair_dangling(env, com, parent, child,
3644                                             comp_id, ea_off, ost_idx, false);
3645
3646         GOTO(log, rc);
3647
3648 log:
3649         if (child != NULL && !IS_ERR(child))
3650                 lfsck_object_put(env, child);
3651
3652         if (parent != NULL && !IS_ERR(parent))
3653                 lfsck_object_put(env, parent);
3654
3655         if (rc)
3656                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant found "
3657                        "dangling reference for: parent "DFID", child "
3658                        DFID", comp_id %u, ea_off %u, ost_idx %u, %s: rc = %d\n",
3659                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
3660                        comp_id, ea_off, ost_idx,
3661                        (lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ) ?
3662                                 "Create the lost OST-object as required" :
3663                                 "Keep the MDT-object there by default", rc);
3664
3665         return rc;
3666 }
3667
3668 /* If the OST-object does not recognize the MDT-object as its parent, and
3669  * there is no other MDT-object claims as its parent, then just trust the
3670  * given MDT-object as its parent. So update the OST-object filter_fid. */
3671 static int lfsck_layout_repair_unmatched_pair(const struct lu_env *env,
3672                                               struct lfsck_component *com,
3673                                               struct dt_object *parent,
3674                                               struct lfsck_layout_req *llr,
3675                                               struct lu_attr *la)
3676 {
3677         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3678         struct filter_fid               *ff     = &info->lti_ff;
3679         struct dt_object                *child  = llr->llr_child;
3680         struct dt_device                *dev    = lfsck_obj2dev(child);
3681         const struct lu_fid             *tfid   = lu_object_fid(&parent->do_lu);
3682         struct lu_buf                   *tbuf   = &info->lti_big_buf;
3683         struct thandle                  *handle;
3684         struct lu_buf                   *buf;
3685         struct lustre_handle             lh     = { 0 };
3686         int                              rc;
3687         ENTRY;
3688
3689         rc = lfsck_ibits_lock(env, com->lc_lfsck, parent, &lh,
3690                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3691                               LCK_EX);
3692         if (rc != 0)
3693                 GOTO(log, rc);
3694
3695         ff->ff_parent.f_seq = cpu_to_le64(tfid->f_seq);
3696         ff->ff_parent.f_oid = cpu_to_le32(tfid->f_oid);
3697         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
3698          * MDT-object's FID::f_ver, instead it is the OST-object index in its
3699          * parent MDT-object's layout EA. */
3700         ff->ff_parent.f_stripe_idx = cpu_to_le32(llr->llr_lov_idx);
3701
3702         rc = lfsck_layout_get_lovea(env, parent, tbuf);
3703         if (unlikely(rc == -ENODATA))
3704                 rc = 0;
3705         if (rc <= 0)
3706                 GOTO(unlock1, rc);
3707
3708         rc = lfsck_lov2layout(tbuf->lb_buf, ff, llr->llr_comp_id);
3709         if (rc)
3710                 GOTO(unlock1, rc);
3711
3712         buf = lfsck_buf_get(env, ff, sizeof(*ff));
3713
3714         handle = dt_trans_create(env, dev);
3715         if (IS_ERR(handle))
3716                 GOTO(unlock1, rc = PTR_ERR(handle));
3717
3718         rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
3719         if (rc != 0)
3720                 GOTO(stop, rc);
3721
3722         rc = dt_attr_get(env, parent, la);
3723         if (rc != 0)
3724                 GOTO(stop, rc);
3725
3726         la->la_valid = LA_UID | LA_GID;
3727         rc = dt_declare_attr_set(env, child, la, handle);
3728         if (rc != 0)
3729                 GOTO(stop, rc);
3730
3731         rc = dt_trans_start_local(env, dev, handle);
3732         if (rc != 0)
3733                 GOTO(stop, rc);
3734
3735         dt_write_lock(env, parent, 0);
3736         if (unlikely(lfsck_is_dead_obj(parent)))
3737                 GOTO(unlock2, rc = 1);
3738
3739         rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
3740         if (rc != 0)
3741                 GOTO(unlock2, rc);
3742
3743         /* Get the latest parent's owner. */
3744         rc = dt_attr_get(env, parent, la);
3745         if (rc != 0)
3746                 GOTO(unlock2, rc);
3747
3748         la->la_valid = LA_UID | LA_GID;
3749         rc = dt_attr_set(env, child, la, handle);
3750
3751         GOTO(unlock2, rc);
3752
3753 unlock2:
3754         dt_write_unlock(env, parent);
3755
3756 stop:
3757         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3758
3759 unlock1:
3760         lfsck_ibits_unlock(&lh, LCK_EX);
3761
3762 log:
3763         if (rc)
3764                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
3765                        "unmatched MDT-OST pair for: parent "DFID
3766                        ", child "DFID", comp_id %u, OST-index %u, "
3767                        "stripe-index %u, owner %u/%u: rc = %d\n",
3768                        lfsck_lfsck2name(com->lc_lfsck),
3769                        PFID(lfsck_dto2fid(parent)),
3770                        PFID(lfsck_dto2fid(child)),
3771                        llr->llr_comp_id, llr->llr_ost_idx, llr->llr_lov_idx,
3772                        la->la_uid, la->la_gid, rc);
3773
3774         return rc;
3775 }
3776
3777 /* If there are more than one MDT-objects claim as the OST-object's parent,
3778  * and the OST-object only recognizes one of them, then we need to generate
3779  * new OST-object(s) with new fid(s) for the non-recognized MDT-object(s). */
3780 static int lfsck_layout_repair_multiple_references(const struct lu_env *env,
3781                                                    struct lfsck_component *com,
3782                                                    struct dt_object *parent,
3783                                                    struct lfsck_layout_req *llr,
3784                                                    struct lu_attr *la)
3785 {
3786         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3787         struct dt_allocation_hint       *hint   = &info->lti_hint;
3788         struct dt_object_format         *dof    = &info->lti_dof;
3789         struct ost_id                   *oi     = &info->lti_oi;
3790         struct lu_buf                   *buf    = &info->lti_big_buf;
3791         struct lfsck_instance           *lfsck  = com->lc_lfsck;
3792         struct dt_device                *dev;
3793         struct lu_device                *d      =
3794                                 &lfsck_obj2dev(llr->llr_child)->dd_lu_dev;
3795         struct lu_object                *o;
3796         struct lu_object                *n;
3797         struct dt_object                *child  = NULL;
3798         struct thandle                  *handle = NULL;
3799         struct lov_mds_md_v1            *lmm;
3800         struct lov_ost_data_v1          *objs;
3801         const struct lu_fid             *pfid   = lfsck_dto2fid(parent);
3802         struct lu_fid                    tfid;
3803         struct lustre_handle             lh     = { 0 };
3804         __u32                            magic;
3805         __u32                            index;
3806         int                              rc;
3807         ENTRY;
3808
3809         /* We use two separated transactions to repair the inconsistency.
3810          *
3811          * 1) create the child (OST-object).
3812          * 2) update the parent LOV EA according to the child's FID.
3813          *
3814          * If 1) succeed, but 2) failed or aborted, then such OST-object will be
3815          * handled as orphan when the layout LFSCK run next time.
3816          *
3817          * If 1) failed, but 2) succeed, then such OST-object will be re-created
3818          * as dangling referened case when the layout LFSCK run next time. */
3819
3820         /* The 1st transaction. */
3821         o = lu_object_anon(env, d, NULL);
3822         if (IS_ERR(o))
3823                 GOTO(log, rc = PTR_ERR(o));
3824
3825         n = lu_object_locate(o->lo_header, d->ld_type);
3826         if (unlikely(n == NULL)) {
3827                 lu_object_put_nocache(env, o);
3828
3829                 GOTO(log, rc = -EINVAL);
3830         }
3831
3832         child = container_of(n, struct dt_object, do_lu);
3833         memset(hint, 0, sizeof(*hint));
3834         rc = dt_attr_get(env, parent, la);
3835         if (rc != 0)
3836                 GOTO(log, rc);
3837
3838         la->la_valid = LA_UID | LA_GID;
3839         memset(dof, 0, sizeof(*dof));
3840
3841         dev = lfsck_obj2dev(child);
3842         handle = dt_trans_create(env, dev);
3843         if (IS_ERR(handle))
3844                 GOTO(log, rc = PTR_ERR(handle));
3845
3846         rc = dt_declare_create(env, child, la, hint, dof, handle);
3847         if (rc != 0)
3848                 GOTO(stop, rc);
3849
3850         rc = dt_trans_start_local(env, dev, handle);
3851         if (rc != 0)
3852                 GOTO(stop, rc);
3853
3854         rc = dt_create(env, child, la, hint, dof, handle);
3855         dt_trans_stop(env, dev, handle);
3856         handle = NULL;
3857         if (rc != 0)
3858                 GOTO(log, rc);
3859
3860         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
3861                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3862                               LCK_EX);
3863         if (rc != 0)
3864                 GOTO(log, rc);
3865
3866         /* The 2nd transaction. */
3867
3868         /* XXX: Generally, we should use bottom device (OSD) to update parent
3869          *      LOV EA. But because the LOD-object still references the wrong
3870          *      OSP-object that should be detached after the parent's LOV EA
3871          *      refreshed. Unfortunately, there is no suitable API for that.
3872          *      So we have to make the LOD to re-load the OSP-object(s) via
3873          *      replacing the LOV EA against the LOD-object.
3874          *
3875          *      Once the DNE2 patches have been landed, we can replace the
3876          *      LOD device with the OSD device. LU-6230. */
3877
3878         dev = lfsck->li_next;
3879         parent = lfsck_object_locate(dev, parent);
3880         if (IS_ERR(parent))
3881                 GOTO(log, rc = PTR_ERR(parent));
3882
3883         handle = dt_trans_create(env, dev);
3884         if (IS_ERR(handle))
3885                 GOTO(log, rc = PTR_ERR(handle));
3886
3887         rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
3888                                   LU_XATTR_REPLACE, handle);
3889         if (rc != 0)
3890                 GOTO(stop, rc);
3891
3892         rc = dt_trans_start_local(env, dev, handle);
3893         if (rc != 0)
3894                 GOTO(stop, rc);
3895
3896         dt_write_lock(env, parent, 0);
3897         if (unlikely(lfsck_is_dead_obj(parent)))
3898                 GOTO(unlock, rc = 0);
3899
3900         rc = lfsck_layout_get_lovea(env, parent, buf);
3901         if (unlikely(rc == -ENODATA))
3902                 rc = 0;
3903         if (rc <= 0)
3904                 GOTO(unlock, rc);
3905
3906         lmm = buf->lb_buf;
3907         magic = le32_to_cpu(lmm->lmm_magic);
3908         if (magic == LOV_MAGIC_COMP_V1) {
3909                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
3910                 struct lov_comp_md_entry_v1 *lcme;
3911                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3912                 int i;
3913
3914                 LASSERT(llr->llr_comp_id != 0);
3915
3916                 for (i = 0; i < count; i++) {
3917                         lcme = &lcm->lcm_entries[i];
3918                         if (le32_to_cpu(lcme->lcme_id) == llr->llr_comp_id) {
3919                                 LASSERT(le32_to_cpu(lcme->lcme_flags) &
3920                                         LCME_FL_INIT);
3921
3922                                 le32_add_cpu(&lcm->lcm_layout_gen, 1);
3923                                 lmm = buf->lb_buf +
3924                                         le32_to_cpu(lcme->lcme_offset);
3925                                 magic = le32_to_cpu(lmm->lmm_magic);
3926                                 goto set;
3927                         }
3928                 }
3929
3930                 GOTO(unlock, rc = 0);
3931         }
3932
3933 set:
3934         if (magic == LOV_MAGIC_V1) {
3935                 objs = &lmm->lmm_objects[llr->llr_lov_idx];
3936         } else {
3937                 LASSERT(magic == LOV_MAGIC_V3);
3938                 objs =
3939                 &((struct lov_mds_md_v3 *)lmm)->lmm_objects[llr->llr_lov_idx];
3940         }
3941
3942         ostid_le_to_cpu(&objs->l_ost_oi, oi);
3943         index = le32_to_cpu(objs->l_ost_idx);
3944         rc = ostid_to_fid(&tfid, oi, index);
3945         /* Someone changed layout during the LFSCK, no need to repair then. */
3946         if (rc == 0 && !lu_fid_eq(&tfid, lu_object_fid(&llr->llr_child->do_lu)))
3947                 GOTO(unlock, rc = 0);
3948
3949         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
3950         fid_to_ostid(lu_object_fid(&child->do_lu), oi);
3951         ostid_cpu_to_le(oi, &objs->l_ost_oi);
3952         objs->l_ost_gen = cpu_to_le32(0);
3953         objs->l_ost_idx = cpu_to_le32(llr->llr_ost_idx);
3954         rc = dt_xattr_set(env, parent, buf, XATTR_NAME_LOV,
3955                           LU_XATTR_REPLACE, handle);
3956
3957         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
3958
3959 unlock:
3960         dt_write_unlock(env, parent);
3961
3962 stop:
3963         if (handle != NULL)
3964                 dt_trans_stop(env, dev, handle);
3965
3966 log:
3967         lfsck_ibits_unlock(&lh, LCK_EX);
3968         if (child != NULL)
3969                 lfsck_object_put(env, child);
3970
3971         if (rc)
3972                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
3973                        "multiple references for: parent "DFID", comp_id %u, "
3974                        "OST-index %u, stripe-index %u, owner %u/%u: rc = %d\n",
3975                        lfsck_lfsck2name(lfsck), PFID(pfid),
3976                        llr->llr_comp_id, llr->llr_ost_idx, llr->llr_lov_idx,
3977                        la->la_uid, la->la_gid, rc);
3978
3979         return rc;
3980 }
3981
3982 /* If the MDT-object and the OST-object have different owner information,
3983  * then trust the MDT-object, because the normal chown/chgrp handle order
3984  * is from MDT to OST, and it is possible that some chown/chgrp operation
3985  * is partly done. */
3986 static int lfsck_layout_repair_owner(const struct lu_env *env,
3987                                      struct lfsck_component *com,
3988                                      struct dt_object *parent,
3989                                      struct lfsck_layout_req *llr,
3990                                      struct lu_attr *pla,
3991                                      const struct lu_attr *cla)
3992 {
3993         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3994         struct lu_attr                  *tla    = &info->lti_la2;
3995         struct dt_object                *child  = llr->llr_child;
3996         struct dt_device                *dev    = lfsck_obj2dev(child);
3997         struct thandle                  *handle;
3998         int                              rc;
3999         dt_obj_version_t                 version;
4000         ENTRY;
4001
4002         tla->la_uid = pla->la_uid;
4003         tla->la_gid = pla->la_gid;
4004         tla->la_valid = LA_UID | LA_GID;
4005         handle = dt_trans_create(env, dev);
4006         if (IS_ERR(handle))
4007                 GOTO(log, rc = PTR_ERR(handle));
4008
4009         rc = dt_declare_attr_set(env, child, tla, handle);
4010         if (rc != 0)
4011                 GOTO(stop, rc);
4012
4013         rc = dt_trans_start_local(env, dev, handle);
4014         if (rc != 0)
4015                 GOTO(stop, rc);
4016
4017         /* Use the dt_object lock to serialize with destroy and attr_set. */
4018         dt_read_lock(env, parent, 0);
4019         if (unlikely(lfsck_is_dead_obj(parent)))
4020                 GOTO(unlock, rc = 1);
4021
4022         version = dt_version_get(env, child);
4023         if (version == -EOPNOTSUPP)
4024                 version = 0;
4025
4026         /* Get the latest parent's owner. */
4027         rc = dt_attr_get(env, parent, pla);
4028         if (rc != 0)
4029                 GOTO(unlock, rc);
4030
4031         /* Some others chown/chgrp during the LFSCK, needs to do nothing. */
4032         if (unlikely((!version && tla->la_ctime == 0) ||
4033                      tla->la_uid != pla->la_uid || tla->la_gid != pla->la_gid))
4034                 rc = 1;
4035         else
4036                 rc = dt_attr_set(env, child, tla, handle);
4037
4038         GOTO(unlock, rc);
4039
4040 unlock:
4041         dt_read_unlock(env, parent);
4042
4043 stop:
4044         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
4045
4046 log:
4047         if (rc != 0)
4048                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
4049                        "inconsistent file owner for: parent "DFID", child "DFID
4050                        ", OST-index %u, stripe-index %u, old owner %u/%u, "
4051                        "new owner %u/%u: rc = %d\n",
4052                        lfsck_lfsck2name(com->lc_lfsck),
4053                        PFID(lfsck_dto2fid(parent)), PFID(lfsck_dto2fid(child)),
4054                        llr->llr_ost_idx, llr->llr_lov_idx,
4055                        cla->la_uid, cla->la_gid, tla->la_uid, tla->la_gid, rc);
4056
4057         return rc;
4058 }
4059
4060 /* Check whether the OST-object correctly back points to the
4061  * MDT-object (@parent) via the XATTR_NAME_FID xattr (@pfid). */
4062 static int lfsck_layout_check_parent(const struct lu_env *env,
4063                                      struct lfsck_component *com,
4064                                      struct lfsck_assistant_object *lso,
4065                                      struct filter_fid *ff,
4066                                      const struct lu_fid *cfid,
4067                                      const struct lu_attr *cla,
4068                                      struct lfsck_layout_req *llr)
4069 {
4070         struct lfsck_thread_info        *info   = lfsck_env_info(env);
4071         struct lu_buf                   *buf    = &info->lti_big_buf;
4072         struct lu_fid                   *pfid   = &info->lti_fid;
4073         struct dt_object                *tobj;
4074         struct lov_mds_md_v1            *lmm;
4075         struct lov_ost_data_v1          *objs;
4076         struct lustre_handle             lh     = { 0 };
4077         int                              rc;
4078         int                              i;
4079         __u32                            magic;
4080         __u32                            idx;
4081         __u16                            count;
4082         ENTRY;
4083
4084         *pfid = ff->ff_parent;
4085         idx = pfid->f_stripe_idx;
4086         pfid->f_ver = 0;
4087
4088         if (unlikely(!fid_is_sane(pfid)))
4089                 RETURN(LLIT_UNMATCHED_PAIR);
4090
4091         if (lu_fid_eq(pfid, &lso->lso_fid)) {
4092                 if (likely(llr->llr_lov_idx == idx))
4093                         RETURN(0);
4094
4095                 RETURN(LLIT_UNMATCHED_PAIR);
4096         }
4097
4098         tobj = lfsck_object_find_bottom(env, com->lc_lfsck, pfid);
4099         if (IS_ERR(tobj))
4100                 RETURN(PTR_ERR(tobj));
4101
4102         if (dt_object_exists(tobj) == 0 || lfsck_is_dead_obj(tobj) ||
4103             !S_ISREG(lfsck_object_type(tobj)))
4104                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4105
4106         /* Load the tobj's layout EA, in spite of it is a local MDT-object or
4107          * remote one on another MDT. Then check whether the given OST-object
4108          * is in such layout. If yes, it is multiple referenced, otherwise it
4109          * is unmatched referenced case. */
4110         rc = lfsck_layout_get_lovea(env, tobj, buf);
4111         if (rc == 0 || rc == -ENODATA || rc == -ENOENT)
4112                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4113
4114         if (unlikely(rc == -EOPNOTSUPP))
4115                 GOTO(out, rc = LLIT_NONE);
4116
4117         if (rc < 0)
4118                 GOTO(out, rc);
4119
4120         lmm = buf->lb_buf;
4121         magic = le32_to_cpu(lmm->lmm_magic);
4122         if (magic == LOV_MAGIC_COMP_V1) {
4123                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
4124                 struct lov_comp_md_entry_v1 *lcme;
4125
4126                 if (ff->ff_layout.ol_comp_id == 0)
4127                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4128
4129                 count = le16_to_cpu(lcm->lcm_entry_count);
4130                 for (i = 0; i < count; i++) {
4131                         lcme = &lcm->lcm_entries[i];
4132                         if (le32_to_cpu(lcme->lcme_id) ==
4133                             ff->ff_layout.ol_comp_id) {
4134                                 lmm = buf->lb_buf +
4135                                         le32_to_cpu(lcme->lcme_offset);
4136                                 magic = le32_to_cpu(lmm->lmm_magic);
4137                                 if (!(le32_to_cpu(lcme->lcme_flags) &
4138                                       LCME_FL_INIT))
4139                                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4140
4141                                 goto further;
4142                         }
4143                 }
4144
4145                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4146         }
4147
4148 further:
4149         if (magic == LOV_MAGIC_V1) {
4150                 objs = &lmm->lmm_objects[0];
4151         } else {
4152                 LASSERT(magic == LOV_MAGIC_V3);
4153                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
4154         }
4155
4156         count = le16_to_cpu(lmm->lmm_stripe_count);
4157         for (i = 0; i < count; i++, objs++) {
4158                 struct lu_fid           *tfid   = &info->lti_fid2;
4159                 struct ost_id           *oi     = &info->lti_oi;
4160                 __u32                    idx2;
4161
4162                 if (lovea_slot_is_dummy(objs))
4163                         continue;
4164
4165                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
4166                 idx2 = le32_to_cpu(objs->l_ost_idx);
4167                 rc = ostid_to_fid(tfid, oi, idx2);
4168                 if (rc != 0) {
4169                         CDEBUG(D_LFSCK, "%s: the parent "DFID" contains "
4170                                "invalid layout EA at the slot %d, index %u\n",
4171                                lfsck_lfsck2name(com->lc_lfsck),
4172                                PFID(pfid), i, idx2);
4173
4174                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4175                 }
4176
4177                 if (lu_fid_eq(cfid, tfid)) {
4178                         rc = lfsck_ibits_lock(env, com->lc_lfsck, tobj, &lh,
4179                                               MDS_INODELOCK_UPDATE |
4180                                               MDS_INODELOCK_LAYOUT |
4181                                               MDS_INODELOCK_XATTR,
4182                                               LCK_EX);
4183                         if (rc != 0)
4184                                 GOTO(out, rc);
4185
4186                         dt_read_lock(env, tobj, 0);
4187
4188                         /* For local MDT-object, re-check existence
4189                          * after taken the lock. */
4190                         if (!dt_object_remote(tobj)) {
4191                                 if (dt_object_exists(tobj) == 0 ||
4192                                     lfsck_is_dead_obj(tobj))
4193                                         rc = LLIT_UNMATCHED_PAIR;
4194                                 else
4195                                         rc = LLIT_MULTIPLE_REFERENCED;
4196
4197                                 GOTO(unlock, rc);
4198                         }
4199
4200                         /* For migration case, the new MDT-object and old
4201                          * MDT-object may reference the same OST-object at
4202                          * some migration internal time.
4203                          *
4204                          * For remote MDT-object, the local MDT may not know
4205                          * whether it has been removed or not.  Try checking
4206                          * for a non-existent xattr to check if this object
4207                          * has been been removed or not. */
4208                         rc = dt_xattr_get(env, tobj, &LU_BUF_NULL,
4209                                           XATTR_NAME_DUMMY);
4210                         if (unlikely(rc == -ENOENT || rc >= 0))
4211                                 rc = LLIT_UNMATCHED_PAIR;
4212                         else if (rc == -ENODATA)
4213                                 rc = LLIT_MULTIPLE_REFERENCED;
4214
4215                         GOTO(unlock, rc);
4216                 }
4217         }
4218
4219         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4220
4221 unlock:
4222         if (lustre_handle_is_used(&lh)) {
4223                 dt_read_unlock(env, tobj);
4224                 lfsck_ibits_unlock(&lh, LCK_EX);
4225         }
4226
4227 out:
4228         lfsck_object_put(env, tobj);
4229
4230         return rc;
4231 }
4232
4233 static int lfsck_layout_assistant_handler_p1(const struct lu_env *env,
4234                                              struct lfsck_component *com,
4235                                              struct lfsck_assistant_req *lar)
4236 {
4237         struct lfsck_layout_req              *llr    =
4238                         container_of0(lar, struct lfsck_layout_req, llr_lar);
4239         struct lfsck_assistant_object        *lso    = lar->lar_parent;
4240         struct lfsck_layout                  *lo     = com->lc_file_ram;
4241         struct lfsck_thread_info             *info   = lfsck_env_info(env);
4242         struct filter_fid                    *ff     = &info->lti_ff;
4243         struct lu_buf buf = { .lb_buf = ff,
4244                               .lb_len = sizeof(*ff) };
4245         struct dt_object                     *parent = NULL;
4246         struct dt_object                     *child  = llr->llr_child;
4247         struct lu_attr                       *pla    = &lso->lso_attr;
4248         struct lu_attr                       *cla    = &info->lti_la;
4249         struct lfsck_instance                *lfsck  = com->lc_lfsck;
4250         struct lfsck_bookmark                *bk     = &lfsck->li_bookmark_ram;
4251         enum lfsck_layout_inconsistency_type  type   = LLIT_NONE;
4252         int                                   rc;
4253         ENTRY;
4254
4255         if (lso->lso_dead)
4256                 RETURN(0);
4257
4258         CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_ENGINE_DELAY, cfs_fail_val);
4259
4260         rc = dt_attr_get(env, child, cla);
4261         if (rc == -ENOENT) {
4262                 parent = lfsck_assistant_object_load(env, lfsck, lso);
4263                 if (IS_ERR(parent)) {
4264                         rc = PTR_ERR(parent);
4265
4266                         RETURN(rc == -ENOENT ? 0 : rc);
4267                 }
4268
4269                 type = LLIT_DANGLING;
4270                 goto repair;
4271         }
4272
4273         if (rc != 0)
4274                 GOTO(out, rc);
4275
4276         lfsck_buf_init(&buf, ff, sizeof(*ff));
4277         rc = dt_xattr_get(env, child, &buf, XATTR_NAME_FID);
4278         if (unlikely(rc > 0 && rc < sizeof(struct lu_fid))) {
4279                 type = LLIT_UNMATCHED_PAIR;
4280                 goto repair;
4281         }
4282
4283         if (rc < 0 && rc != -ENODATA)
4284                 GOTO(out, rc);
4285
4286         if (rc == 0 || rc == -ENODATA)
4287                 GOTO(check_owner, rc = 0);
4288
4289         filter_fid_le_to_cpu(ff, ff, sizeof(*ff));
4290         rc = lfsck_layout_check_parent(env, com, lso, ff,
4291                                        lu_object_fid(&child->do_lu), cla, llr);
4292         if (rc > 0) {
4293                 type = rc;
4294                 goto repair;
4295         }
4296
4297         if (rc < 0)
4298                 GOTO(out, rc);
4299
4300 check_owner:
4301         /* Someone may has changed the owner after the parent attr pre-loaded.
4302          * It can be handled later inside the lfsck_layout_repair_owner(). */
4303         if (unlikely(cla->la_uid != pla->la_uid ||
4304                      cla->la_gid != pla->la_gid)) {
4305                 type = LLIT_INCONSISTENT_OWNER;
4306                 goto repair;
4307         }
4308
4309 repair:
4310         if (type == LLIT_NONE)
4311                 GOTO(out, rc = 0);
4312
4313         if (bk->lb_param & LPF_DRYRUN)
4314                 GOTO(out, rc = 1);
4315
4316         if (parent == NULL) {
4317                 parent = lfsck_assistant_object_load(env, lfsck, lso);
4318                 if (IS_ERR(parent)) {
4319                         rc = PTR_ERR(parent);
4320
4321                         if (rc == -ENOENT)
4322                                 RETURN(0);
4323
4324                         GOTO(out, rc);
4325                 }
4326         }
4327
4328         switch (type) {
4329         case LLIT_DANGLING:
4330                 if (bk->lb_param & LPF_DELAY_CREATE_OSTOBJ)
4331                         rc = lfsck_layout_ins_dangling_rec(env, com,
4332                                 lfsck_dto2fid(parent), lfsck_dto2fid(child),
4333                                 llr->llr_comp_id, llr->llr_lov_idx,
4334                                 llr->llr_ost_idx);
4335                 else
4336                         rc = __lfsck_layout_repair_dangling(env, com, parent,
4337                                                             llr->llr_child,
4338                                                             llr->llr_comp_id,
4339                                                             llr->llr_lov_idx,
4340                                                             llr->llr_ost_idx,
4341                                                             true);
4342                 break;
4343         case LLIT_UNMATCHED_PAIR:
4344                 rc = lfsck_layout_repair_unmatched_pair(env, com, parent,
4345                                                         llr, pla);
4346                 break;
4347         case LLIT_MULTIPLE_REFERENCED:
4348                 rc = lfsck_layout_repair_multiple_references(env, com, parent,
4349                                                              llr, pla);
4350                 break;
4351         case LLIT_INCONSISTENT_OWNER:
4352                 rc = lfsck_layout_repair_owner(env, com, parent, llr, pla, cla);
4353                 break;
4354         default:
4355                 rc = 0;
4356                 break;
4357         }
4358
4359         GOTO(out, rc);
4360
4361 out:
4362         down_write(&com->lc_sem);
4363         if (rc < 0) {
4364                 struct lfsck_assistant_data *lad = com->lc_data;
4365
4366                 if (unlikely(test_bit(LAD_EXIT, &lad->lad_flags))) {
4367                         rc = 0;
4368                 } else if (rc == -ENOTCONN || rc == -ESHUTDOWN ||
4369                            rc == -ETIMEDOUT || rc == -EHOSTDOWN ||
4370                            rc == -EHOSTUNREACH) {
4371                         /* If cannot touch the target server,
4372                          * mark the LFSCK as INCOMPLETE. */
4373                         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant fail to "
4374                                "talk with OST %x: rc = %d\n",
4375                                lfsck_lfsck2name(lfsck), llr->llr_ost_idx, rc);
4376                         lfsck_lad_set_bitmap(env, com, llr->llr_ost_idx);
4377                         lo->ll_objs_skipped++;
4378                         rc = 0;
4379                 } else {
4380                         lfsck_layout_record_failure(env, lfsck, lo);
4381                 }
4382         } else if (rc > 0 && (type != LLIT_DANGLING ||
4383                               !(bk->lb_param & LPF_DELAY_CREATE_OSTOBJ))) {
4384                 LASSERTF(type > LLIT_NONE && type <= LLIT_MAX,
4385                          "unknown type = %d\n", type);
4386
4387                 lo->ll_objs_repaired[type - 1]++;
4388                 if (bk->lb_param & LPF_DRYRUN &&
4389                     unlikely(lo->ll_pos_first_inconsistent == 0))
4390                         lo->ll_pos_first_inconsistent =
4391                         lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
4392                                                         lfsck->li_di_oit);
4393         }
4394         up_write(&com->lc_sem);
4395
4396         if (parent != NULL && !IS_ERR(parent))
4397                 lfsck_object_put(env, parent);
4398
4399         return rc;
4400 }
4401
4402 static int
4403 lfsck_layout_double_scan_one_trace_file(const struct lu_env *env,
4404                                         struct lfsck_component *com,
4405                                         struct dt_object *obj, bool first)
4406 {
4407         struct lfsck_instance *lfsck = com->lc_lfsck;
4408         struct ptlrpc_thread *thread = &lfsck->li_thread;
4409         struct lfsck_bookmark *bk = &lfsck->li_bookmark_ram;
4410         struct lfsck_layout *lo = com->lc_file_ram;
4411         const struct dt_it_ops *iops = &obj->do_index_ops->dio_it;
4412         struct dt_it *di;
4413         struct dt_key *key;
4414         struct lfsck_layout_dangling_key *parent =
4415                                         &lfsck_env_info(env)->lti_lldk;
4416         struct lu_fid *cfid = &lfsck_env_info(env)->lti_fid3;
4417         __u32 ost_idx;
4418         int rc;
4419         ENTRY;
4420
4421         di = iops->init(env, obj, 0);
4422         if (IS_ERR(di))
4423                 RETURN(PTR_ERR(di));
4424
4425         if (first)
4426                 lldk_cpu_to_be(parent, &lo->ll_lldk_latest_scanned_phase2);
4427         else
4428                 memset(parent, 0, sizeof(*parent));
4429         rc = iops->get(env, di, (const struct dt_key *)parent);
4430         if (rc < 0)
4431                 GOTO(fini, rc);
4432
4433         if (first) {
4434                 /* The start one either has been processed or does not exist,
4435                  * skip it. */
4436                 rc = iops->next(env, di);
4437                 if (rc != 0)
4438                         GOTO(put, rc);
4439         }
4440
4441         do {
4442                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
4443                     unlikely(!thread_is_running(thread)))
4444                         GOTO(put, rc = 0);
4445
4446                 key = iops->key(env, di);
4447                 if (IS_ERR(key)) {
4448                         rc = PTR_ERR(key);
4449                         if (rc == -ENOENT)
4450                                 GOTO(put, rc = 1);
4451
4452                         goto checkpoint;
4453                 }
4454
4455                 lldk_be_to_cpu(parent,
4456                                 (const struct lfsck_layout_dangling_key *)key);
4457                 if (!fid_is_sane(&parent->lldk_fid)) {
4458                         rc = 0;
4459                         goto checkpoint;
4460                 }
4461
4462                 rc = iops->rec(env, di, (struct dt_rec *)cfid, 0);
4463                 if (rc == 0) {
4464                         fid_be_to_cpu(cfid, cfid);
4465                         ost_idx = cfid->f_ver;
4466                         cfid->f_ver = 0;
4467                         if (!fid_is_sane(cfid)) {
4468                                 rc = 0;
4469                                 goto checkpoint;
4470                         }
4471
4472                         rc = lfsck_layout_repair_dangling(env, com,
4473                                         &parent->lldk_fid, cfid,
4474                                         parent->lldk_comp_id,
4475                                         parent->lldk_ea_off, ost_idx);
4476                 }
4477
4478 checkpoint:
4479                 down_write(&com->lc_sem);
4480                 com->lc_new_checked++;
4481                 com->lc_new_scanned++;
4482                 if (rc >= 0)
4483                         lo->ll_lldk_latest_scanned_phase2 = *parent;
4484
4485                 if (rc > 0)
4486                         lo->ll_objs_repaired[LLIT_DANGLING - 1]++;
4487                 else if (rc < 0)
4488                         lo->ll_objs_failed_phase2++;
4489                 up_write(&com->lc_sem);
4490
4491                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
4492                         GOTO(put, rc);
4493
4494                 if (unlikely(com->lc_time_next_checkpoint <=
4495                              ktime_get_seconds()) &&
4496                     com->lc_new_checked != 0) {
4497                         down_write(&com->lc_sem);
4498                         lo->ll_run_time_phase2 += ktime_get_seconds() -
4499                                                   com->lc_time_last_checkpoint;
4500                         lo->ll_time_last_checkpoint = ktime_get_real_seconds();
4501                         lo->ll_objs_checked_phase2 += com->lc_new_checked;
4502                         com->lc_new_checked = 0;
4503                         lfsck_layout_store(env, com);
4504                         up_write(&com->lc_sem);
4505
4506                         com->lc_time_last_checkpoint = ktime_get_seconds();
4507                         com->lc_time_next_checkpoint =
4508                                 com->lc_time_last_checkpoint +
4509                                 LFSCK_CHECKPOINT_INTERVAL;
4510                 }
4511
4512                 lfsck_control_speed_by_self(com);
4513                 if (unlikely(!thread_is_running(thread)))
4514                         GOTO(put, rc = 0);
4515
4516                 rc = iops->next(env, di);
4517         } while (rc == 0);
4518
4519         GOTO(put, rc);
4520
4521 put:
4522         iops->put(env, di);
4523
4524 fini:
4525         iops->fini(env, di);
4526
4527         return rc;
4528 }
4529
4530 static int lfsck_layout_assistant_handler_p2(const struct lu_env *env,
4531                                              struct lfsck_component *com)
4532 {
4533         struct lfsck_assistant_data     *lad    = com->lc_data;
4534         struct lfsck_instance           *lfsck  = com->lc_lfsck;
4535         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
4536         struct lfsck_tgt_descs          *ltds   = &lfsck->li_ost_descs;
4537         struct lfsck_tgt_desc           *ltd;
4538         int                              rc     = 0;
4539         ENTRY;
4540
4541         CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan start\n",
4542                lfsck_lfsck2name(lfsck));
4543
4544         spin_lock(&ltds->ltd_lock);
4545         while (!list_empty(&lad->lad_ost_phase2_list)) {
4546                 ltd = list_entry(lad->lad_ost_phase2_list.next,
4547                                  struct lfsck_tgt_desc,
4548                                  ltd_layout_phase_list);
4549                 list_del_init(&ltd->ltd_layout_phase_list);
4550                 if (bk->lb_param & LPF_OST_ORPHAN) {
4551                         spin_unlock(&ltds->ltd_lock);
4552                         rc = lfsck_layout_scan_orphan(env, com, ltd);
4553                         if (rc != 0 && bk->lb_param & LPF_FAILOUT)
4554                                 RETURN(rc);
4555
4556                         if (unlikely(test_bit(LAD_EXIT, &lad->lad_flags) ||
4557                                      !thread_is_running(&lfsck->li_thread)))
4558                                 RETURN(0);
4559                         spin_lock(&ltds->ltd_lock);
4560                 }
4561         }
4562
4563         if (list_empty(&lad->lad_ost_phase1_list))
4564                 rc = 1;
4565         else
4566                 rc = 0;
4567         spin_unlock(&ltds->ltd_lock);
4568
4569         if (rc == 1 && bk->lb_param & LPF_OST_ORPHAN) {
4570                 struct lfsck_layout *lo = com->lc_file_ram;
4571                 int i;
4572
4573                 com->lc_new_checked = 0;
4574                 com->lc_new_scanned = 0;
4575                 com->lc_time_last_checkpoint = ktime_get_seconds();
4576                 com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
4577                                                LFSCK_CHECKPOINT_INTERVAL;
4578
4579                 i = lfsck_sub_trace_file_fid2idx(
4580                                 &lo->ll_lldk_latest_scanned_phase2.lldk_fid);
4581                 rc = lfsck_layout_double_scan_one_trace_file(env, com,
4582                                 com->lc_sub_trace_objs[i].lsto_obj, true);
4583                 while (rc > 0 && ++i < LFSCK_STF_COUNT)
4584                         rc = lfsck_layout_double_scan_one_trace_file(env, com,
4585                                 com->lc_sub_trace_objs[i].lsto_obj, false);
4586
4587                 CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan dangling stop "
4588                        "at the No. %d trace file: rc = %d\n",
4589                        lfsck_lfsck2name(lfsck), i, rc);
4590         }
4591
4592         CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan stop: rc = %d\n",
4593                lfsck_lfsck2name(lfsck), rc);
4594
4595         RETURN(rc);
4596 }
4597
4598 static int
4599 lfsck_layout_slave_async_interpret(const struct lu_env *env,
4600                                    struct ptlrpc_request *req,
4601                                    void *args, int rc)
4602 {
4603         struct lfsck_layout_slave_async_args *llsaa = args;
4604         struct obd_export *exp = llsaa->llsaa_exp;
4605         struct lfsck_component *com = llsaa->llsaa_com;
4606         struct lfsck_layout_slave_target *llst = llsaa->llsaa_llst;
4607         struct lfsck_layout_slave_data *llsd = com->lc_data;
4608         struct lfsck_reply *lr = NULL;
4609         bool done = false;
4610
4611         if (rc != 0) {
4612                 /* It is probably caused by network trouble, or target crash,
4613                  * it will try several times (depends on the obd_timeout, and
4614                  * will not less than 3 times). But to make the LFSCK can go
4615                  * ahead, we should not try for ever. After some try but still
4616                  * hit failure, it will assume that the target exit the LFSCK
4617                  * prcoessing and stop try. */
4618                 if (rc == -ENOTCONN || rc == -ESHUTDOWN) {
4619                         int max_try = max_t(int, obd_timeout / 30, 3);
4620
4621                         if (++(llst->llst_failures) > max_try)
4622                                 done = true;
4623                 } else {
4624                         done = true;
4625                 }
4626         } else {
4627                 llst->llst_failures = 0;
4628                 lr = req_capsule_server_get(&req->rq_pill, &RMF_LFSCK_REPLY);
4629                 if (lr->lr_status != LS_SCANNING_PHASE1 &&
4630                     lr->lr_status != LS_SCANNING_PHASE2)
4631                         done = true;
4632         }
4633
4634         if (done) {
4635                 CDEBUG(D_LFSCK, "%s: layout LFSCK slave gets the MDT %x "
4636                        "status %d, failures_try %d\n", lfsck_lfsck2name(com->lc_lfsck),
4637                        llst->llst_index, lr != NULL ? lr->lr_status : rc,
4638                        llst->llst_failures);
4639
4640                 lfsck_layout_llst_del(llsd, llst);
4641         }
4642
4643         lfsck_layout_llst_put(llst);
4644         lfsck_component_put(env, com);
4645         class_export_put(exp);
4646
4647         return 0;
4648 }
4649
4650 static int lfsck_layout_async_query(const struct lu_env *env,
4651                                     struct lfsck_component *com,
4652                                     struct obd_export *exp,
4653                                     struct lfsck_layout_slave_target *llst,
4654                                     struct lfsck_request *lr,
4655                                     struct ptlrpc_request_set *set)
4656 {
4657         struct lfsck_layout_slave_async_args *llsaa;
4658         struct ptlrpc_request                *req;
4659         struct lfsck_request                 *tmp;
4660         int                                   rc;
4661         ENTRY;
4662
4663         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_QUERY);
4664         if (req == NULL)
4665                 RETURN(-ENOMEM);
4666
4667         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_QUERY);
4668         if (rc != 0) {
4669                 ptlrpc_request_free(req);
4670                 RETURN(rc);
4671         }
4672
4673         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4674         *tmp = *lr;
4675         ptlrpc_request_set_replen(req);
4676
4677         llsaa = ptlrpc_req_async_args(llsaa, req);
4678         llsaa->llsaa_exp = exp;
4679         llsaa->llsaa_com = lfsck_component_get(com);
4680         llsaa->llsaa_llst = llst;
4681         req->rq_interpret_reply = lfsck_layout_slave_async_interpret;
4682         req->rq_allow_intr = 1;
4683         req->rq_no_delay = 1;
4684         ptlrpc_set_add_req(set, req);
4685
4686         RETURN(0);
4687 }
4688
4689 static int lfsck_layout_async_notify(const struct lu_env *env,
4690                                      struct obd_export *exp,
4691                                      struct lfsck_request *lr,
4692                                      struct ptlrpc_request_set *set)
4693 {
4694         struct ptlrpc_request   *req;
4695         struct lfsck_request    *tmp;
4696         int                      rc;
4697         ENTRY;
4698
4699         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
4700         if (req == NULL)
4701                 RETURN(-ENOMEM);
4702
4703         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
4704         if (rc != 0) {
4705                 ptlrpc_request_free(req);
4706                 RETURN(rc);
4707         }
4708
4709         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4710         *tmp = *lr;
4711         ptlrpc_request_set_replen(req);
4712         req->rq_allow_intr = 1;
4713         req->rq_no_delay = 1;
4714         ptlrpc_set_add_req(set, req);
4715
4716         RETURN(0);
4717 }
4718
4719 static int
4720 lfsck_layout_slave_query_master(const struct lu_env *env,
4721                                 struct lfsck_component *com)
4722 {
4723         struct lfsck_request             *lr    = &lfsck_env_info(env)->lti_lr;
4724         struct lfsck_instance            *lfsck = com->lc_lfsck;
4725         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
4726         struct lfsck_layout_slave_target *llst;
4727         struct obd_export                *exp;
4728         struct ptlrpc_request_set        *set;
4729         int                               rc    = 0;
4730         int                               rc1   = 0;
4731         ENTRY;
4732
4733         set = ptlrpc_prep_set();
4734         if (set == NULL)
4735                 GOTO(log, rc = -ENOMEM);
4736
4737         memset(lr, 0, sizeof(*lr));
4738         lr->lr_event = LE_QUERY;
4739         lr->lr_active = LFSCK_TYPE_LAYOUT;
4740
4741         llsd->llsd_touch_gen++;
4742         spin_lock(&llsd->llsd_lock);
4743         while (!list_empty(&llsd->llsd_master_list)) {
4744                 llst = list_entry(llsd->llsd_master_list.next,
4745                                   struct lfsck_layout_slave_target,
4746                                   llst_list);
4747                 if (llst->llst_gen == llsd->llsd_touch_gen)
4748                         break;
4749
4750                 llst->llst_gen = llsd->llsd_touch_gen;
4751                 list_move_tail(&llst->llst_list,
4752                                &llsd->llsd_master_list);
4753                 atomic_inc(&llst->llst_ref);
4754                 spin_unlock(&llsd->llsd_lock);
4755
4756                 exp = lustre_find_lwp_by_index(lfsck->li_obd->obd_name,
4757                                                llst->llst_index);
4758                 if (exp == NULL) {
4759                         lfsck_layout_llst_del(llsd, llst);
4760                         lfsck_layout_llst_put(llst);
4761                         spin_lock(&llsd->llsd_lock);
4762                         continue;
4763                 }
4764
4765                 rc = lfsck_layout_async_query(env, com, exp, llst, lr, set);
4766                 if (rc != 0) {
4767                         CDEBUG(D_LFSCK, "%s: layout LFSCK slave fail to "
4768                                "query %s for layout: rc = %d\n",
4769                                lfsck_lfsck2name(lfsck),
4770                                exp->exp_obd->obd_name, rc);
4771
4772                         rc1 = rc;
4773                         lfsck_layout_llst_put(llst);
4774                         class_export_put(exp);
4775                 }
4776                 spin_lock(&llsd->llsd_lock);
4777         }
4778         spin_unlock(&llsd->llsd_lock);
4779
4780         rc = ptlrpc_set_wait(env, set);
4781         ptlrpc_set_destroy(set);
4782
4783         GOTO(log, rc = (rc1 != 0 ? rc1 : rc));
4784
4785 log:
4786         CDEBUG(D_LFSCK, "%s: layout LFSCK slave queries master: rc = %d\n",
4787                lfsck_lfsck2name(com->lc_lfsck), rc);
4788
4789         return rc;
4790 }
4791
4792 static void
4793 lfsck_layout_slave_notify_master(const struct lu_env *env,
4794                                  struct lfsck_component *com,
4795                                  enum lfsck_events event, int result)
4796 {
4797         struct lfsck_layout              *lo    = com->lc_file_ram;
4798         struct lfsck_instance            *lfsck = com->lc_lfsck;
4799         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
4800         struct lfsck_request             *lr    = &lfsck_env_info(env)->lti_lr;
4801         struct lfsck_layout_slave_target *llst;
4802         struct obd_export                *exp;
4803         struct ptlrpc_request_set        *set;
4804         int                               rc;
4805         ENTRY;
4806
4807         CDEBUG(D_LFSCK, "%s: layout LFSCK slave notifies master\n",
4808                lfsck_lfsck2name(com->lc_lfsck));
4809
4810         set = ptlrpc_prep_set();
4811         if (set == NULL)
4812                 RETURN_EXIT;
4813
4814         memset(lr, 0, sizeof(*lr));
4815         lr->lr_event = event;
4816         lr->lr_flags = LEF_FROM_OST;
4817         lr->lr_status = result;
4818         lr->lr_index = lfsck_dev_idx(lfsck);
4819         lr->lr_active = LFSCK_TYPE_LAYOUT;
4820         lr->lr_flags2 = lo->ll_flags;
4821         llsd->llsd_touch_gen++;
4822         spin_lock(&llsd->llsd_lock);
4823         while (!list_empty(&llsd->llsd_master_list)) {
4824                 llst = list_entry(llsd->llsd_master_list.next,
4825                                   struct lfsck_layout_slave_target,
4826                                   llst_list);
4827                 if (llst->llst_gen == llsd->llsd_touch_gen)
4828                         break;
4829
4830                 llst->llst_gen = llsd->llsd_touch_gen;
4831                 list_move_tail(&llst->llst_list,
4832                                &llsd->llsd_master_list);
4833                 atomic_inc(&llst->llst_ref);
4834                 spin_unlock(&llsd->llsd_lock);
4835
4836                 exp = lustre_find_lwp_by_index(lfsck->li_obd->obd_name,
4837                                                llst->llst_index);
4838                 if (exp == NULL) {
4839                         lfsck_layout_llst_del(llsd, llst);
4840                         lfsck_layout_llst_put(llst);
4841                         spin_lock(&llsd->llsd_lock);
4842                         continue;
4843                 }
4844
4845                 rc = lfsck_layout_async_notify(env, exp, lr, set);
4846                 if (rc != 0)
4847                         CDEBUG(D_LFSCK, "%s: layout LFSCK slave fail to "
4848                                "notify %s for layout: rc = %d\n",
4849                                lfsck_lfsck2name(lfsck),
4850                                exp->exp_obd->obd_name, rc);
4851
4852                 lfsck_layout_llst_put(llst);
4853                 class_export_put(exp);
4854                 spin_lock(&llsd->llsd_lock);
4855         }
4856         spin_unlock(&llsd->llsd_lock);
4857
4858         ptlrpc_set_wait(env, set);
4859         ptlrpc_set_destroy(set);
4860
4861         RETURN_EXIT;
4862 }
4863
4864 /*
4865  * \ret -ENODATA: unrecognized stripe
4866  * \ret = 0     : recognized stripe
4867  * \ret < 0     : other failures
4868  */
4869 static int lfsck_layout_master_check_pairs(const struct lu_env *env,
4870                                            struct lfsck_component *com,
4871                                            struct lu_fid *cfid,
4872                                            struct lu_fid *pfid, __u32 comp_id)
4873 {
4874         struct lfsck_thread_info        *info   = lfsck_env_info(env);
4875         struct lu_buf                   *buf    = &info->lti_big_buf;
4876         struct ost_id                   *oi     = &info->lti_oi;
4877         struct dt_object                *obj;
4878         struct lov_mds_md_v1            *lmm;
4879         struct lov_ost_data_v1          *objs;
4880         __u32                            idx    = pfid->f_stripe_idx;
4881         __u32                            magic;
4882         int                              rc     = 0;
4883         int                              i;
4884         __u16                            count;
4885         ENTRY;
4886
4887         pfid->f_ver = 0;
4888         obj = lfsck_object_find_bottom(env, com->lc_lfsck, pfid);
4889         if (IS_ERR(obj))
4890                 RETURN(PTR_ERR(obj));
4891
4892         dt_read_lock(env, obj, 0);
4893         if (unlikely(dt_object_exists(obj) == 0 ||
4894                      lfsck_is_dead_obj(obj)))
4895                 GOTO(unlock, rc = -ENOENT);
4896
4897         if (!S_ISREG(lfsck_object_type(obj)))
4898                 GOTO(unlock, rc = -ENODATA);
4899
4900         rc = lfsck_layout_get_lovea(env, obj, buf);
4901         if (rc < 0)
4902                 GOTO(unlock, rc);
4903
4904         lmm = buf->lb_buf;
4905         magic = le32_to_cpu(lmm->lmm_magic);
4906         if (magic == LOV_MAGIC_COMP_V1) {
4907                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
4908                 struct lov_comp_md_entry_v1 *lcme;
4909
4910                 if (comp_id == 0)
4911                         GOTO(unlock, rc = -ENODATA);
4912
4913                 count = le16_to_cpu(lcm->lcm_entry_count);
4914                 for (i = 0; i < count; i++) {
4915                         lcme = &lcm->lcm_entries[i];
4916                         if (le32_to_cpu(lcme->lcme_id) == comp_id) {
4917                                 lmm = buf->lb_buf +
4918                                         le32_to_cpu(lcme->lcme_offset);
4919                                 magic = le32_to_cpu(lmm->lmm_magic);
4920                                 if (!(le32_to_cpu(lcme->lcme_flags) &
4921                                       LCME_FL_INIT))
4922                                         GOTO(unlock, rc = -ENODATA);
4923
4924                                 goto further;
4925                         }
4926                 }
4927
4928                 GOTO(unlock, rc = -ENODATA);
4929         }
4930
4931 further:
4932         if (magic == LOV_MAGIC_V1) {
4933                 objs = &lmm->lmm_objects[0];
4934         } else {
4935                 LASSERT(magic == LOV_MAGIC_V3);
4936                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
4937         }
4938
4939         fid_to_ostid(cfid, oi);
4940         count = le16_to_cpu(lmm->lmm_stripe_count);
4941         for (i = 0; i < count; i++, objs++) {
4942                 struct ost_id oi2;
4943
4944                 ostid_le_to_cpu(&objs->l_ost_oi, &oi2);
4945                 if (memcmp(oi, &oi2, sizeof(*oi)) == 0)
4946                         GOTO(unlock, rc = (i != idx ? -ENODATA : 0));
4947         }
4948
4949         GOTO(unlock, rc = -ENODATA);
4950
4951 unlock:
4952         dt_read_unlock(env, obj);
4953         lfsck_object_put(env, obj);
4954
4955         return rc;
4956 }
4957
4958 /*
4959  * The LFSCK-on-OST will ask the LFSCK-on-MDT to check whether the given
4960  * MDT-object/OST-object pairs match or not to aviod transfer MDT-object
4961  * layout EA from MDT to OST. On one hand, the OST no need to understand
4962  * the layout EA structure; on the other hand, it may cause trouble when
4963  * transfer large layout EA from MDT to OST via normal OUT RPC.
4964  *
4965  * \ret > 0: unrecognized stripe
4966  * \ret = 0: recognized stripe
4967  * \ret < 0: other failures
4968  */
4969 static int lfsck_layout_slave_check_pairs(const struct lu_env *env,
4970                                           struct lfsck_component *com,
4971                                           struct lu_fid *cfid,
4972                                           struct lu_fid *pfid, __u32 comp_id)
4973 {
4974         struct lfsck_instance    *lfsck  = com->lc_lfsck;
4975         struct obd_device        *obd    = lfsck->li_obd;
4976         struct seq_server_site   *ss     = lfsck_dev_site(lfsck);
4977         struct obd_export        *exp    = NULL;
4978         struct ptlrpc_request    *req    = NULL;
4979         struct lfsck_request     *lr;
4980         struct lu_seq_range      *range  = &lfsck_env_info(env)->lti_range;
4981         int                       rc     = 0;
4982         ENTRY;
4983
4984         if (unlikely(fid_is_idif(pfid)))
4985                 RETURN(1);
4986
4987         fld_range_set_any(range);
4988         rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(pfid), range);
4989         if (rc != 0)
4990                 RETURN(rc == -ENOENT ? 1 : rc);
4991
4992         if (unlikely(!fld_range_is_mdt(range)))
4993                 RETURN(1);
4994
4995         exp = lustre_find_lwp_by_index(obd->obd_name, range->lsr_index);
4996         if (unlikely(exp == NULL))
4997                 RETURN(1);
4998
4999         if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
5000                 GOTO(out, rc = -EOPNOTSUPP);
5001
5002         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
5003         if (req == NULL)
5004                 GOTO(out, rc = -ENOMEM);
5005
5006         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
5007         if (rc != 0) {
5008                 ptlrpc_request_free(req);
5009
5010                 GOTO(out, rc);
5011         }
5012
5013         lr = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
5014         memset(lr, 0, sizeof(*lr));
5015         lr->lr_event = LE_PAIRS_VERIFY;
5016         lr->lr_active = LFSCK_TYPE_LAYOUT;
5017         lr->lr_fid = *cfid; /* OST-object itself FID. */
5018         lr->lr_fid2 = *pfid; /* The claimed parent FID. */
5019         lr->lr_comp_id = comp_id;
5020
5021         ptlrpc_request_set_replen(req);
5022         rc = ptlrpc_queue_wait(req);
5023         ptlrpc_req_finished(req);
5024
5025         if (rc == -ENOENT || rc == -ENODATA)
5026                 rc = 1;
5027
5028         GOTO(out, rc);
5029
5030 out:
5031         if (exp != NULL)
5032                 class_export_put(exp);
5033
5034         return rc;
5035 }
5036
5037 static int lfsck_layout_slave_repair_pfid(const struct lu_env *env,
5038                                           struct lfsck_component *com,
5039                                           struct lfsck_req_local *lrl)
5040 {
5041         struct dt_object        *obj;
5042         int                      rc     = 0;
5043         ENTRY;
5044
5045         obj = lfsck_object_find_bottom(env, com->lc_lfsck, &lrl->lrl_fid);
5046         if (IS_ERR(obj))
5047                 GOTO(log, rc = PTR_ERR(obj));
5048
5049         dt_write_lock(env, obj, 0);
5050         if (unlikely(dt_object_exists(obj) == 0 ||
5051                      lfsck_is_dead_obj(obj)))
5052                 GOTO(unlock, rc = 0);
5053
5054         rc = __lfsck_layout_update_pfid(env, obj, &lrl->lrl_ff_client.ff_parent,
5055                                         &lrl->lrl_ff_client.ff_layout,
5056                                         lrl->lrl_ff_client.ff_layout_version,
5057                                         lrl->lrl_ff_client.ff_range,
5058                                         lrl->lrl_ff_client.ff_parent.f_ver);
5059
5060         GOTO(unlock, rc);
5061
5062 unlock:
5063         dt_write_unlock(env, obj);
5064         lfsck_object_put(env, obj);
5065
5066 log:
5067         CDEBUG(D_LFSCK, "%s: layout LFSCK slave repaired pfid for "DFID
5068                ", parent "DFID": rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
5069                PFID(&lrl->lrl_fid), PFID(&lrl->lrl_ff_client.ff_parent), rc);
5070
5071         return rc;
5072 }
5073
5074 /* layout APIs */
5075
5076 static void lfsck_layout_slave_quit(const struct lu_env *env,
5077                                     struct lfsck_component *com);
5078
5079 static int lfsck_layout_reset(const struct lu_env *env,
5080                               struct lfsck_component *com, bool init)
5081 {
5082         struct lfsck_layout     *lo    = com->lc_file_ram;
5083         int                      rc;
5084
5085         down_write(&com->lc_sem);
5086         if (init) {
5087                 memset(lo, 0, com->lc_file_size);
5088         } else {
5089                 __u32 count = lo->ll_success_count;
5090                 time64_t last_time = lo->ll_time_last_complete;
5091
5092                 memset(lo, 0, com->lc_file_size);
5093                 lo->ll_success_count = count;
5094                 lo->ll_time_last_complete = last_time;
5095         }
5096
5097         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
5098         lo->ll_status = LS_INIT;
5099
5100         if (com->lc_lfsck->li_master) {
5101                 struct lfsck_assistant_data *lad = com->lc_data;
5102
5103                 clear_bit(LAD_INCOMPLETE, &lad->lad_flags);
5104                 CFS_RESET_BITMAP(lad->lad_bitmap);
5105         }
5106
5107         rc = lfsck_layout_store(env, com);
5108         if (rc == 0 && com->lc_lfsck->li_master)
5109                 rc = lfsck_load_sub_trace_files(env, com,
5110                         &dt_lfsck_layout_dangling_features, LFSCK_LAYOUT, true);
5111         up_write(&com->lc_sem);
5112
5113         CDEBUG(D_LFSCK, "%s: layout LFSCK reset: rc = %d\n",
5114                lfsck_lfsck2name(com->lc_lfsck), rc);
5115
5116         return rc;
5117 }
5118
5119 static void lfsck_layout_fail(const struct lu_env *env,
5120                               struct lfsck_component *com, bool new_checked)
5121 {
5122         struct lfsck_layout *lo = com->lc_file_ram;
5123
5124         down_write(&com->lc_sem);
5125         if (new_checked)
5126                 com->lc_new_checked++;
5127         lfsck_layout_record_failure(env, com->lc_lfsck, lo);
5128         up_write(&com->lc_sem);
5129 }
5130
5131 static int lfsck_layout_master_checkpoint(const struct lu_env *env,
5132                                           struct lfsck_component *com, bool init)
5133 {
5134         struct lfsck_instance   *lfsck   = com->lc_lfsck;
5135         struct lfsck_layout     *lo      = com->lc_file_ram;
5136         int                      rc;
5137
5138         if (!init) {
5139                 rc = lfsck_checkpoint_generic(env, com);
5140                 if (rc != 0)
5141                         return rc > 0 ? 0 : rc;
5142         }
5143
5144         down_write(&com->lc_sem);
5145         if (init) {
5146                 lo->ll_pos_latest_start =
5147                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5148         } else {
5149                 lo->ll_pos_last_checkpoint =
5150                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5151                 lo->ll_run_time_phase1 += ktime_get_seconds() -
5152                                           lfsck->li_time_last_checkpoint;
5153                 lo->ll_time_last_checkpoint = ktime_get_real_seconds();
5154                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5155                 com->lc_new_checked = 0;
5156         }
5157
5158         rc = lfsck_layout_store(env, com);
5159         up_write(&com->lc_sem);
5160
5161         CDEBUG(D_LFSCK, "%s: layout LFSCK master checkpoint at the pos ["
5162                "%llu], status = %d: rc = %d\n", lfsck_lfsck2name(lfsck),
5163                lfsck->li_pos_current.lp_oit_cookie, lo->ll_status, rc);
5164
5165         return rc;
5166 }
5167
5168 static int lfsck_layout_slave_checkpoint(const struct lu_env *env,
5169                                          struct lfsck_component *com, bool init)
5170 {
5171         struct lfsck_instance   *lfsck = com->lc_lfsck;
5172         struct lfsck_layout     *lo    = com->lc_file_ram;
5173         int                      rc;
5174
5175         if (com->lc_new_checked == 0 && !init)
5176                 return 0;
5177
5178         down_write(&com->lc_sem);
5179         if (init) {
5180                 lo->ll_pos_latest_start =
5181                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5182         } else {
5183                 lo->ll_pos_last_checkpoint =
5184                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5185                 lo->ll_run_time_phase1 += ktime_get_seconds() -
5186                                           lfsck->li_time_last_checkpoint;
5187                 lo->ll_time_last_checkpoint = ktime_get_real_seconds();
5188                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5189                 com->lc_new_checked = 0;
5190         }
5191
5192         rc = lfsck_layout_store(env, com);
5193         up_write(&com->lc_sem);
5194
5195         CDEBUG(D_LFSCK, "%s: layout LFSCK slave checkpoint at the pos ["
5196                "%llu], status = %d: rc = %d\n", lfsck_lfsck2name(lfsck),
5197                lfsck->li_pos_current.lp_oit_cookie, lo->ll_status, rc);
5198
5199         return rc;
5200 }
5201
5202 static int lfsck_layout_prep(const struct lu_env *env,
5203                              struct lfsck_component *com,
5204                              struct lfsck_start *start)
5205 {
5206         struct lfsck_instance   *lfsck  = com->lc_lfsck;
5207         struct lfsck_layout     *lo     = com->lc_file_ram;
5208         struct lfsck_position   *pos    = &com->lc_pos_start;
5209
5210         fid_zero(&pos->lp_dir_parent);
5211         pos->lp_dir_cookie = 0;
5212         if (lo->ll_status == LS_COMPLETED ||
5213             lo->ll_status == LS_PARTIAL ||
5214             /* To handle orphan, must scan from the beginning. */
5215             (start != NULL && start->ls_flags & LPF_OST_ORPHAN)) {
5216                 int rc;
5217
5218                 rc = lfsck_layout_reset(env, com, false);
5219                 if (rc == 0)
5220                         rc = lfsck_set_param(env, lfsck, start, true);
5221
5222                 if (rc != 0) {
5223                         CDEBUG(D_LFSCK, "%s: layout LFSCK prep failed: "
5224                                "rc = %d\n", lfsck_lfsck2name(lfsck), rc);
5225
5226                         return rc;
5227                 }
5228         }
5229
5230         down_write(&com->lc_sem);
5231         lo->ll_time_latest_start = ktime_get_real_seconds();
5232         spin_lock(&lfsck->li_lock);
5233         if (lo->ll_flags & LF_SCANNED_ONCE) {
5234                 if (!lfsck->li_drop_dryrun ||
5235                     lo->ll_pos_first_inconsistent == 0) {
5236                         lo->ll_status = LS_SCANNING_PHASE2;
5237                         list_move_tail(&com->lc_link,
5238                                        &lfsck->li_list_double_scan);
5239                         pos->lp_oit_cookie = 0;
5240                 } else {
5241                         int i;
5242
5243                         lo->ll_status = LS_SCANNING_PHASE1;
5244                         lo->ll_run_time_phase1 = 0;
5245                         lo->ll_run_time_phase2 = 0;
5246                         lo->ll_objs_checked_phase1 = 0;
5247                         lo->ll_objs_checked_phase2 = 0;
5248                         lo->ll_objs_failed_phase1 = 0;
5249                         lo->ll_objs_failed_phase2 = 0;
5250                         for (i = 0; i < LLIT_MAX; i++)
5251                                 lo->ll_objs_repaired[i] = 0;
5252
5253                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
5254                         fid_zero(&com->lc_fid_latest_scanned_phase2);
5255                 }
5256         } else {
5257                 lo->ll_status = LS_SCANNING_PHASE1;
5258                 if (!lfsck->li_drop_dryrun ||
5259                     lo->ll_pos_first_inconsistent == 0)
5260                         pos->lp_oit_cookie = lo->ll_pos_last_checkpoint + 1;
5261                 else
5262                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
5263         }
5264         spin_unlock(&lfsck->li_lock);
5265         up_write(&com->lc_sem);
5266
5267         return 0;
5268 }
5269
5270 static int lfsck_layout_slave_prep(const struct lu_env *env,
5271                                    struct lfsck_component *com,
5272                                    struct lfsck_start_param *lsp)
5273 {
5274         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
5275         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5276         struct lfsck_layout             *lo     = com->lc_file_ram;
5277         struct lfsck_start              *start  = lsp->lsp_start;
5278         int                              rc;
5279
5280         rc = lfsck_layout_prep(env, com, start);
5281         if (rc != 0)
5282                 return rc;
5283
5284         if (lo->ll_flags & LF_CRASHED_LASTID &&
5285             list_empty(&llsd->llsd_master_list)) {
5286                 LASSERT(lfsck->li_out_notify != NULL);
5287
5288                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5289                                      LE_LASTID_REBUILDING);
5290         }
5291
5292         if (!lsp->lsp_index_valid)
5293                 return 0;
5294
5295         rc = lfsck_layout_llst_add(llsd, lsp->lsp_index);
5296         if (rc == 0 && start != NULL && start->ls_flags & LPF_OST_ORPHAN) {
5297                 LASSERT(!llsd->llsd_rbtree_valid);
5298
5299                 down_write(&llsd->llsd_rb_rwsem);
5300                 rc = lfsck_rbtree_setup(env, com);
5301                 up_write(&llsd->llsd_rb_rwsem);
5302         }
5303
5304         CDEBUG(D_LFSCK, "%s: layout LFSCK slave prep done, start pos ["
5305                "%llu]\n", lfsck_lfsck2name(lfsck),
5306                com->lc_pos_start.lp_oit_cookie);
5307
5308         return rc;
5309 }
5310
5311 static int lfsck_layout_master_prep(const struct lu_env *env,
5312                                     struct lfsck_component *com,
5313                                     struct lfsck_start_param *lsp)
5314 {
5315         int rc;
5316         ENTRY;
5317
5318         rc = lfsck_layout_load_bitmap(env, com);
5319         if (rc != 0) {
5320                 rc = lfsck_layout_reset(env, com, false);
5321                 if (rc == 0)
5322                         rc = lfsck_set_param(env, com->lc_lfsck,
5323                                              lsp->lsp_start, true);
5324
5325                 if (rc != 0)
5326                         GOTO(log, rc);
5327         }
5328
5329         rc = lfsck_layout_prep(env, com, lsp->lsp_start);
5330         if (rc != 0)
5331                 RETURN(rc);
5332
5333         rc = lfsck_start_assistant(env, com, lsp);
5334
5335         GOTO(log, rc);
5336
5337 log:
5338         CDEBUG(D_LFSCK, "%s: layout LFSCK master prep done, start pos ["
5339                "%llu]\n", lfsck_lfsck2name(com->lc_lfsck),
5340                com->lc_pos_start.lp_oit_cookie);
5341
5342         return 0;
5343 }
5344
5345 /* Pre-fetch the attribute for each stripe in the given layout EA. */
5346 static int lfsck_layout_scan_stripes(const struct lu_env *env,
5347                                      struct lfsck_component *com,
5348                                      struct dt_object *parent,
5349                                      struct lov_mds_md_v1 *lmm, __u32 comp_id)
5350 {
5351         struct lfsck_thread_info        *info    = lfsck_env_info(env);
5352         struct lfsck_instance           *lfsck   = com->lc_lfsck;
5353         struct lfsck_bookmark           *bk      = &lfsck->li_bookmark_ram;
5354         struct lfsck_layout             *lo      = com->lc_file_ram;
5355         struct lfsck_assistant_data     *lad     = com->lc_data;
5356         struct lfsck_assistant_object   *lso     = NULL;
5357         struct lov_ost_data_v1          *objs;
5358         struct lfsck_tgt_descs          *ltds    = &lfsck->li_ost_descs;
5359         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
5360         struct ptlrpc_thread            *athread = &lad->lad_thread;
5361         struct lu_buf                    buf;
5362         int                              rc      = 0;
5363         int                              i;
5364         __u32                            magic;
5365         __u16                            count;
5366         ENTRY;
5367
5368         lfsck_buf_init(&buf, &info->lti_ff, sizeof(struct filter_fid));
5369         magic = le32_to_cpu(lmm->lmm_magic);
5370         if (magic == LOV_MAGIC_V1) {
5371                 objs = &lmm->lmm_objects[0];
5372         } else {
5373                 LASSERT(magic == LOV_MAGIC_V3);
5374                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
5375         }
5376
5377         count = le16_to_cpu(lmm->lmm_stripe_count);
5378         for (i = 0; i < count; i++, objs++) {
5379                 struct lu_fid           *fid    = &info->lti_fid;
5380                 struct ost_id           *oi     = &info->lti_oi;
5381                 struct lfsck_layout_req *llr;
5382                 struct lfsck_tgt_desc   *tgt    = NULL;
5383                 struct dt_object        *cobj   = NULL;
5384                 __u32                    index;
5385                 bool                     wakeup = false;
5386
5387                 if (unlikely(lovea_slot_is_dummy(objs)))
5388                         continue;
5389
5390                 wait_event_idle(mthread->t_ctl_waitq,
5391                                 lad->lad_prefetched < bk->lb_async_windows ||
5392                                 !thread_is_running(mthread) ||
5393                                 thread_is_stopped(athread));
5394
5395                 if (unlikely(!thread_is_running(mthread)) ||
5396                              thread_is_stopped(athread))
5397                         GOTO(out, rc = 0);
5398
5399                 if (unlikely(lfsck_is_dead_obj(parent)))
5400                         GOTO(out, rc = 0);
5401
5402                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
5403                 index = le32_to_cpu(objs->l_ost_idx);
5404                 rc = ostid_to_fid(fid, oi, index);
5405                 if (rc != 0) {
5406                         CDEBUG(D_LFSCK, "%s: get invalid layout EA for "DFID
5407                                ": "DOSTID", idx %u, comp_id %u\n",
5408                                lfsck_lfsck2name(lfsck),
5409                                PFID(lfsck_dto2fid(parent)), POSTID(oi),
5410                                index, comp_id);
5411                         goto next;
5412                 }
5413
5414                 tgt = lfsck_tgt_get(ltds, index);
5415                 if (unlikely(tgt == NULL)) {
5416                         CDEBUG(D_LFSCK, "%s: cannot talk with OST %x which "
5417                                "did not join the layout LFSCK, comp_id %u\n",
5418                                lfsck_lfsck2name(lfsck), index, comp_id);
5419                         lfsck_lad_set_bitmap(env, com, index);
5420                         goto next;
5421                 }
5422
5423                 /* There is potential deadlock race condition between object
5424                  * destroy and layout LFSCK. Consider the following scenario:
5425                  *
5426                  * 1) The LFSCK thread obtained the parent object firstly, at
5427                  *    that time, the parent object has not been destroyed yet.
5428                  *
5429                  * 2) One RPC service thread destroyed the parent and all its
5430                  *    children objects. Because the LFSCK is referencing the
5431                  *    parent object, then the parent object will be marked as
5432                  *    dying in RAM. On the other hand, the parent object is
5433                  *    referencing all its children objects, then all children
5434                  *    objects will be marked as dying in RAM also.
5435                  *
5436                  * 3) The LFSCK thread tries to find some child object with
5437                  *    the parent object referenced. Then it will find that the
5438                  *    child object is dying. According to the object visibility
5439                  *    rules: the object with dying flag cannot be returned to
5440                  *    others. So the LFSCK thread has to wait until the dying
5441                  *    object has been purged from RAM, then it can allocate a
5442                  *    new object (with the same FID) in RAM. Unfortunately, the
5443                  *    LFSCK thread itself is referencing the parent object, and
5444                  *    cause the parent object cannot be purged, then cause the
5445                  *    child object cannot be purged also. So the LFSCK thread
5446                  *    will fall into deadlock.
5447                  */
5448                 cobj = lfsck_object_find_by_dev(env, tgt->ltd_tgt, fid);
5449                 if (IS_ERR(cobj)) {
5450                         if (lfsck_is_dead_obj(parent)) {
5451                                 lfsck_tgt_put(tgt);
5452
5453                                 GOTO(out, rc = 0);
5454                         }
5455
5456                         rc = PTR_ERR(cobj);
5457                         goto next;
5458                 }
5459
5460                 rc = dt_declare_attr_get(env, cobj);
5461                 if (rc)
5462                         goto next;
5463
5464                 rc = dt_declare_xattr_get(env, cobj, &buf, XATTR_NAME_FID);
5465                 if (rc)
5466                         goto next;
5467
5468                 if (lso == NULL) {
5469                         struct lu_attr *attr = &info->lti_la;
5470
5471                         rc = dt_attr_get(env, parent, attr);
5472                         if (rc != 0)
5473                                 goto next;
5474
5475                         lso = lfsck_assistant_object_init(env,
5476                                 lfsck_dto2fid(parent), attr,
5477                                 lfsck->li_pos_current.lp_oit_cookie, false);
5478                         if (IS_ERR(lso)) {
5479                                 rc = PTR_ERR(lso);
5480                                 lso = NULL;
5481
5482                                 goto next;
5483                         }
5484                 }
5485
5486                 llr = lfsck_layout_assistant_req_init(lso, cobj, comp_id,
5487                                                       index, i);
5488                 if (IS_ERR(llr)) {
5489                         rc = PTR_ERR(llr);
5490                         goto next;
5491                 }
5492
5493                 cobj = NULL;
5494                 spin_lock(&lad->lad_lock);
5495                 if (lad->lad_assistant_status < 0) {
5496                         spin_unlock(&lad->lad_lock);
5497                         lfsck_layout_assistant_req_fini(env, &llr->llr_lar);
5498                         lfsck_tgt_put(tgt);
5499                         RETURN(lad->lad_assistant_status);
5500                 }
5501
5502                 list_add_tail(&llr->llr_lar.lar_list, &lad->lad_req_list);
5503                 if (lad->lad_prefetched == 0)
5504                         wakeup = true;
5505
5506                 lad->lad_prefetched++;
5507                 spin_unlock(&lad->lad_lock);
5508                 if (wakeup)
5509                         wake_up_all(&athread->t_ctl_waitq);
5510
5511 next:
5512                 down_write(&com->lc_sem);
5513                 com->lc_new_checked++;
5514                 if (rc < 0)
5515                         lfsck_layout_record_failure(env, lfsck, lo);
5516                 up_write(&com->lc_sem);
5517
5518                 if (cobj != NULL && !IS_ERR(cobj))
5519                         lfsck_object_put(env, cobj);
5520
5521                 if (likely(tgt != NULL))
5522                         lfsck_tgt_put(tgt);
5523
5524                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
5525                         GOTO(out, rc);
5526         }
5527
5528         GOTO(out, rc = 0);
5529
5530 out:
5531         if (lso != NULL)
5532                 lfsck_assistant_object_put(env, lso);
5533
5534         return rc;
5535 }
5536
5537 /* For the given object, read its layout EA locally. For each stripe, pre-fetch
5538  * the OST-object's attribute and generate an structure lfsck_layout_req on the
5539  * list ::lad_req_list.
5540  *
5541  * For each request on above list, the lfsck_layout_assistant thread compares
5542  * the OST side attribute with local attribute, if inconsistent, then repair it.
5543  *
5544  * All above processing is async mode with pipeline. */
5545 static int lfsck_layout_master_exec_oit(const struct lu_env *env,
5546                                         struct lfsck_component *com,
5547                                         struct dt_object *obj)
5548 {
5549         struct lfsck_thread_info        *info   = lfsck_env_info(env);
5550         struct ost_id                   *oi     = &info->lti_oi;
5551         struct lfsck_layout             *lo     = com->lc_file_ram;
5552         struct lfsck_assistant_data     *lad    = com->lc_data;
5553         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5554         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
5555         struct thandle                  *handle = NULL;
5556         struct lu_buf                   *buf    = &info->lti_big_buf;
5557         struct lov_mds_md_v1            *lmm    = NULL;
5558         struct dt_device                *dev    = lfsck_obj2dev(obj);
5559         struct lustre_handle             lh     = { 0 };
5560         struct lu_buf                    ea_buf = { NULL };
5561         struct lov_comp_md_v1           *lcm    = NULL;
5562         struct lov_comp_md_entry_v1     *lcme   = NULL;
5563         int                              rc     = 0;
5564         int                              size   = 0;
5565         __u32                            magic  = 0;
5566         __u16                            count  = 0;
5567         bool                             locked = false;
5568         bool                             stripe = false;
5569         bool                             bad_oi = false;
5570         ENTRY;
5571
5572         if (!S_ISREG(lfsck_object_type(obj)))
5573                 GOTO(out, rc = 0);
5574
5575         if (lad->lad_assistant_status < 0)
5576                 GOTO(out, rc = -ESRCH);
5577
5578         fid_to_lmm_oi(lfsck_dto2fid(obj), oi);
5579         lmm_oi_cpu_to_le(oi, oi);
5580         dt_read_lock(env, obj, 0);
5581         locked = true;
5582
5583 again:
5584         bad_oi = false;
5585         if (dt_object_exists(obj) == 0 ||
5586             lfsck_is_dead_obj(obj))
5587                 GOTO(out, rc = 0);
5588
5589         rc = lfsck_layout_get_lovea(env, obj, buf);
5590         if (rc == -EINVAL || rc == -ENODATA || rc == -EOPNOTSUPP)
5591                 /* Skip bad lov EA during the 1st cycle scanning, and
5592                  * try to recover it via orphan in the 2nd scanning. */
5593                 rc = 0;
5594         if (rc <= 0)
5595                 GOTO(out, rc);
5596
5597         size = rc;
5598         lmm = buf->lb_buf;
5599         magic = le32_to_cpu(lmm->lmm_magic);
5600         if (magic == LOV_MAGIC_COMP_V1) {
5601                 int i;
5602
5603                 lcm = buf->lb_buf;
5604                 count = le16_to_cpu(lcm->lcm_entry_count);
5605                 for (i = 0; i < count; i++) {
5606                         lcme = &lcm->lcm_entries[i];
5607                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
5608                         if (memcmp(oi, &lmm->lmm_oi, sizeof(*oi)) != 0)
5609                                 goto fix;
5610                 }
5611
5612                 GOTO(out, stripe = true);
5613         } else if (memcmp(oi, &lmm->lmm_oi, sizeof(*oi)) == 0) {
5614                 GOTO(out, stripe = true);
5615         }
5616
5617 fix:
5618         /* Inconsistent lmm_oi, should be repaired. */
5619         bad_oi = true;
5620
5621         if (bk->lb_param & LPF_DRYRUN) {
5622                 lo->ll_objs_repaired[LLIT_OTHERS - 1]++;
5623
5624                 GOTO(out, stripe = true);
5625         }
5626
5627         if (!lustre_handle_is_used(&lh)) {
5628                 dt_read_unlock(env, obj);
5629                 locked = false;
5630                 rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
5631                                       MDS_INODELOCK_LAYOUT |
5632                                       MDS_INODELOCK_XATTR, LCK_EX);
5633                 if (rc != 0)
5634                         GOTO(out, rc);
5635
5636                 handle = dt_trans_create(env, dev);
5637                 if (IS_ERR(handle))
5638                         GOTO(out, rc = PTR_ERR(handle));
5639
5640                 lfsck_buf_init(&ea_buf, buf->lb_buf, size);
5641                 rc = dt_declare_xattr_set(env, obj, &ea_buf, XATTR_NAME_LOV,
5642                                           LU_XATTR_REPLACE, handle);
5643                 if (rc != 0)
5644                         GOTO(out, rc);
5645
5646                 rc = dt_trans_start_local(env, dev, handle);
5647                 if (rc != 0)
5648                         GOTO(out, rc);
5649
5650                 dt_write_lock(env, obj, 0);
5651                 locked = true;
5652
5653                 goto again;
5654         }
5655
5656         if (magic == LOV_MAGIC_COMP_V1) {
5657                 int i;
5658
5659                 for (i = 0; i < count; i++) {
5660                         lcme = &lcm->lcm_entries[i];
5661                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
5662                         lmm->lmm_oi = *oi;
5663                 }
5664         } else {
5665                 lmm->lmm_oi = *oi;
5666         }
5667
5668         rc = dt_xattr_set(env, obj, &ea_buf, XATTR_NAME_LOV,
5669                           LU_XATTR_REPLACE, handle);
5670         if (rc != 0)
5671                 GOTO(out, rc);
5672
5673         lo->ll_objs_repaired[LLIT_OTHERS - 1]++;
5674
5675         GOTO(out, stripe = true);
5676
5677 out:
5678         if (locked) {
5679                 if (lustre_handle_is_used(&lh))
5680                         dt_write_unlock(env, obj);
5681                 else
5682                         dt_read_unlock(env, obj);
5683         }
5684
5685         if (handle != NULL && !IS_ERR(handle))
5686                 dt_trans_stop(env, dev, handle);
5687
5688         lfsck_ibits_unlock(&lh, LCK_EX);
5689
5690         if (bad_oi)
5691                 CDEBUG(D_LFSCK, "%s: layout LFSCK master %s bad lmm_oi for "
5692                        DFID": rc = %d\n", lfsck_lfsck2name(lfsck),
5693                        bk->lb_param & LPF_DRYRUN ? "found" : "repaired",
5694                        PFID(lfsck_dto2fid(obj)), rc);
5695
5696         if (stripe) {
5697                 if (magic == LOV_MAGIC_COMP_V1) {
5698                         int i;
5699
5700                         for (i = 0; i < count; i++) {
5701                                 lcme = &lcm->lcm_entries[i];
5702                                 if (!(le32_to_cpu(lcme->lcme_flags) &
5703                                       LCME_FL_INIT))
5704                                         continue;
5705
5706                                 rc = lfsck_layout_scan_stripes(env, com, obj,
5707                                         (struct lov_mds_md_v1 *)(buf->lb_buf +
5708                                         le32_to_cpu(lcme->lcme_offset)),
5709                                         le32_to_cpu(lcme->lcme_id));
5710                         }
5711                 } else {
5712                         rc = lfsck_layout_scan_stripes(env, com, obj, lmm, 0);
5713                 }
5714         } else {
5715                 down_write(&com->lc_sem);
5716                 com->lc_new_checked++;
5717                 if (rc < 0)
5718                         lfsck_layout_record_failure(env, lfsck, lo);
5719                 up_write(&com->lc_sem);
5720         }
5721
5722         return rc;
5723 }
5724
5725 static int lfsck_layout_slave_exec_oit(const struct lu_env *env,
5726                                        struct lfsck_component *com,
5727                                        struct dt_object *obj)
5728 {
5729         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5730         struct lfsck_layout             *lo     = com->lc_file_ram;
5731         const struct lu_fid             *fid    = lfsck_dto2fid(obj);
5732         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
5733         struct lfsck_layout_seq         *lls;
5734         __u64                            seq;
5735         __u64                            oid;
5736         int                              rc;
5737         ENTRY;
5738
5739         LASSERT(llsd != NULL);
5740
5741         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY5) &&
5742             cfs_fail_val == lfsck_dev_idx(lfsck)) {
5743                 struct ptlrpc_thread    *thread = &lfsck->li_thread;
5744
5745                 wait_event_idle_timeout(thread->t_ctl_waitq,
5746                                         !thread_is_running(thread),
5747                                         cfs_time_seconds(1));
5748         }
5749
5750         lfsck_rbtree_update_bitmap(env, com, fid, false);
5751
5752         down_write(&com->lc_sem);
5753         if (fid_is_idif(fid))
5754                 seq = 0;
5755         else if (!fid_is_norm(fid) ||
5756                  !fid_is_for_ostobj(env, lfsck, obj, fid))
5757                 GOTO(unlock, rc = 0);
5758         else
5759                 seq = fid_seq(fid);
5760         com->lc_new_checked++;
5761
5762         lls = lfsck_layout_seq_lookup(llsd, seq);
5763         if (lls == NULL) {
5764                 OBD_ALLOC_PTR(lls);
5765                 if (unlikely(lls == NULL))
5766                         GOTO(unlock, rc = -ENOMEM);
5767
5768                 INIT_LIST_HEAD(&lls->lls_list);
5769                 lls->lls_seq = seq;
5770                 rc = lfsck_layout_lastid_load(env, com, lls);
5771                 if (rc != 0) {
5772                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to "
5773                               "load LAST_ID for %#llx: rc = %d\n",
5774                               lfsck_lfsck2name(com->lc_lfsck), seq, rc);
5775                         lo->ll_objs_failed_phase1++;
5776                         OBD_FREE_PTR(lls);
5777                         GOTO(unlock, rc);
5778                 }
5779
5780                 lfsck_layout_seq_insert(llsd, lls);
5781         }
5782
5783         if (unlikely(fid_is_last_id(fid)))
5784                 GOTO(unlock, rc = 0);
5785
5786         if (fid_is_idif(fid))
5787                 oid = fid_idif_id(fid_seq(fid), fid_oid(fid), fid_ver(fid));
5788         else
5789                 oid = fid_oid(fid);
5790
5791         if (oid > lls->lls_lastid_known)
5792                 lls->lls_lastid_known = oid;
5793
5794         if (oid > lls->lls_lastid) {
5795                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
5796                         /* OFD may create new objects during LFSCK scanning. */
5797                         rc = lfsck_layout_lastid_reload(env, com, lls);
5798                         if (unlikely(rc != 0)) {
5799                                 CDEBUG(D_LFSCK, "%s: layout LFSCK failed to "
5800                                       "reload LAST_ID for %#llx: rc = %d\n",
5801                                       lfsck_lfsck2name(com->lc_lfsck),
5802                                       lls->lls_seq, rc);
5803
5804                                 GOTO(unlock, rc);
5805                         }
5806
5807                         if (oid <= lls->lls_lastid ||
5808                             lo->ll_flags & LF_CRASHED_LASTID)
5809                                 GOTO(unlock, rc = 0);
5810
5811                         LASSERT(lfsck->li_out_notify != NULL);
5812
5813                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5814                                              LE_LASTID_REBUILDING);
5815                         lo->ll_flags |= LF_CRASHED_LASTID;
5816
5817                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds crashed "
5818                                "LAST_ID file (2) for the sequence %#llx"
5819                                ", old value %llu, known value %llu\n",
5820                                lfsck_lfsck2name(lfsck), lls->lls_seq,
5821                                lls->lls_lastid, oid);
5822                 }
5823
5824                 lls->lls_lastid = oid;
5825                 lls->lls_dirty = 1;
5826         }
5827
5828         GOTO(unlock, rc = 0);
5829
5830 unlock:
5831         up_write(&com->lc_sem);
5832
5833         return rc;
5834 }
5835
5836 static int lfsck_layout_exec_dir(const struct lu_env *env,
5837                                  struct lfsck_component *com,
5838                                  struct lfsck_assistant_object *lso,
5839                                  struct lu_dirent *ent, __u16 type)
5840 {
5841         return 0;
5842 }
5843
5844 static int lfsck_layout_master_post(const struct lu_env *env,
5845                                     struct lfsck_component *com,
5846                                     int result, bool init)
5847 {
5848         struct lfsck_instance   *lfsck  = com->lc_lfsck;
5849         struct lfsck_layout     *lo     = com->lc_file_ram;
5850         int                      rc;
5851         ENTRY;
5852
5853         lfsck_post_generic(env, com, &result);
5854
5855         down_write(&com->lc_sem);
5856         spin_lock(&lfsck->li_lock);
5857         if (!init)
5858                 lo->ll_pos_last_checkpoint =
5859                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5860
5861         if (result > 0) {
5862                 if (lo->ll_flags & LF_INCOMPLETE)
5863                         lo->ll_status = LS_PARTIAL;
5864                 else
5865                         lo->ll_status = LS_SCANNING_PHASE2;
5866                 lo->ll_flags |= LF_SCANNED_ONCE;
5867                 lo->ll_flags &= ~LF_UPGRADE;
5868                 list_move_tail(&com->lc_link, &lfsck->li_list_double_scan);
5869         } else if (result == 0) {
5870                 if (lfsck->li_status != 0)
5871                         lo->ll_status = lfsck->li_status;
5872                 else
5873                         lo->ll_status = LS_STOPPED;
5874                 if (lo->ll_status != LS_PAUSED)
5875                         list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5876         } else {
5877                 lo->ll_status = LS_FAILED;
5878                 list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5879         }
5880         spin_unlock(&lfsck->li_lock);
5881
5882         if (!init) {
5883                 lo->ll_run_time_phase1 += ktime_get_seconds() -
5884                                           lfsck->li_time_last_checkpoint;
5885                 lo->ll_time_last_checkpoint = ktime_get_real_seconds();
5886                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5887                 com->lc_new_checked = 0;
5888         }
5889
5890         rc = lfsck_layout_store(env, com);
5891         up_write(&com->lc_sem);
5892
5893         CDEBUG(D_LFSCK, "%s: layout LFSCK master post done: rc = %d\n",
5894                lfsck_lfsck2name(lfsck), rc);
5895
5896         RETURN(rc);
5897 }
5898
5899 static int lfsck_layout_slave_post(const struct lu_env *env,
5900                                    struct lfsck_component *com,
5901                                    int result, bool init)
5902 {
5903         struct lfsck_instance   *lfsck = com->lc_lfsck;
5904         struct lfsck_layout     *lo    = com->lc_file_ram;
5905         int                      rc;
5906         bool                     done  = false;
5907
5908         down_write(&com->lc_sem);
5909         rc = lfsck_layout_lastid_store(env, com);
5910         if (rc != 0)
5911                 result = rc;
5912
5913         LASSERT(lfsck->li_out_notify != NULL);
5914
5915         spin_lock(&lfsck->li_lock);
5916         if (!init)
5917                 lo->ll_pos_last_checkpoint =
5918                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5919
5920         if (result > 0) {
5921                 lo->ll_status = LS_SCANNING_PHASE2;
5922                 lo->ll_flags |= LF_SCANNED_ONCE;
5923                 if (lo->ll_flags & LF_CRASHED_LASTID) {
5924                         done = true;
5925                         lo->ll_flags &= ~LF_CRASHED_LASTID;
5926
5927                         CDEBUG(D_LFSCK, "%s: layout LFSCK has rebuilt "
5928                                "crashed LAST_ID files successfully\n",
5929                                lfsck_lfsck2name(lfsck));
5930                 }
5931                 lo->ll_flags &= ~LF_UPGRADE;
5932                 list_move_tail(&com->lc_link, &lfsck->li_list_double_scan);
5933         } else if (result == 0) {
5934                 if (lfsck->li_status != 0)
5935                         lo->ll_status = lfsck->li_status;
5936                 else
5937                         lo->ll_status = LS_STOPPED;
5938                 if (lo->ll_status != LS_PAUSED)
5939                         list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5940         } else {
5941                 lo->ll_status = LS_FAILED;
5942                 list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5943         }
5944         spin_unlock(&lfsck->li_lock);
5945
5946         if (done)
5947                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5948                                      LE_LASTID_REBUILT);
5949
5950         if (!init) {
5951                 lo->ll_run_time_phase1 += ktime_get_seconds() -
5952                                           lfsck->li_time_last_checkpoint;
5953                 lo->ll_time_last_checkpoint = ktime_get_real_seconds();
5954                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5955                 com->lc_new_checked = 0;
5956         }
5957
5958         rc = lfsck_layout_store(env, com);
5959         up_write(&com->lc_sem);
5960
5961         lfsck_layout_slave_notify_master(env, com, LE_PHASE1_DONE, result);
5962
5963         CDEBUG(D_LFSCK, "%s: layout LFSCK slave post done: rc = %d\n",
5964                lfsck_lfsck2name(lfsck), rc);
5965
5966         return rc;
5967 }
5968
5969 static void lfsck_layout_dump(const struct lu_env *env,
5970                               struct lfsck_component *com, struct seq_file *m)
5971 {
5972         struct lfsck_instance   *lfsck = com->lc_lfsck;
5973         struct lfsck_bookmark   *bk    = &lfsck->li_bookmark_ram;
5974         struct lfsck_layout     *lo    = com->lc_file_ram;
5975         const char *prefix;
5976
5977         down_read(&com->lc_sem);
5978         if (bk->lb_param & LPF_DRYRUN)
5979                 prefix = "inconsistent";
5980         else
5981                 prefix = "repaired";
5982
5983         seq_printf(m, "name: lfsck_layout\n"
5984                    "magic: %#x\n"
5985                    "version: %d\n"
5986                    "status: %s\n",
5987                    lo->ll_magic,
5988                    bk->lb_version,
5989                    lfsck_status2name(lo->ll_status));
5990
5991         lfsck_bits_dump(m, lo->ll_flags, lfsck_flags_names, "flags");
5992
5993         lfsck_bits_dump(m, bk->lb_param, lfsck_param_names, "param");
5994
5995         lfsck_time_dump(m, lo->ll_time_last_complete, "last_completed");
5996
5997         lfsck_time_dump(m, lo->ll_time_latest_start, "latest_start");
5998
5999         lfsck_time_dump(m, lo->ll_time_last_checkpoint, "last_checkpoint");
6000
6001         seq_printf(m, "latest_start_position: %llu\n"
6002                    "last_checkpoint_position: %llu\n"
6003                    "first_failure_position: %llu\n",
6004                    lo->ll_pos_latest_start,
6005                    lo->ll_pos_last_checkpoint,
6006                    lo->ll_pos_first_inconsistent);
6007
6008         seq_printf(m, "success_count: %u\n"
6009                    "%s_dangling: %llu\n"
6010                    "%s_unmatched_pair: %llu\n"
6011                    "%s_multiple_referenced: %llu\n"
6012                    "%s_orphan: %llu\n"
6013                    "%s_inconsistent_owner: %llu\n"
6014                    "%s_others: %llu\n"
6015                    "skipped: %llu\n"
6016                    "failed_phase1: %llu\n"
6017                    "failed_phase2: %llu\n",
6018                    lo->ll_success_count,
6019                    prefix, lo->ll_objs_repaired[LLIT_DANGLING - 1],
6020                    prefix, lo->ll_objs_repaired[LLIT_UNMATCHED_PAIR - 1],
6021                    prefix, lo->ll_objs_repaired[LLIT_MULTIPLE_REFERENCED - 1],
6022                    prefix, lo->ll_objs_repaired[LLIT_ORPHAN - 1],
6023                    prefix, lo->ll_objs_repaired[LLIT_INCONSISTENT_OWNER - 1],
6024                    prefix, lo->ll_objs_repaired[LLIT_OTHERS - 1],
6025                    lo->ll_objs_skipped,
6026                    lo->ll_objs_failed_phase1,
6027                    lo->ll_objs_failed_phase2);
6028
6029         if (lo->ll_status == LS_SCANNING_PHASE1) {
6030                 time64_t duration = ktime_get_seconds() -
6031                                     lfsck->li_time_last_checkpoint;
6032                 u64 checked = lo->ll_objs_checked_phase1 +
6033                               com->lc_new_checked;
6034                 u64 speed = checked;
6035                 u64 new_checked = com->lc_new_checked;
6036                 time64_t rtime = lo->ll_run_time_phase1 + duration;
6037                 u64 pos;
6038
6039                 if (duration != 0)
6040                         new_checked = div64_s64(new_checked, duration);
6041                 if (rtime != 0)
6042                         speed = div64_s64(speed, rtime);
6043                 seq_printf(m, "checked_phase1: %llu\n"
6044                            "checked_phase2: %llu\n"
6045                            "run_time_phase1: %lld seconds\n"
6046                            "run_time_phase2: %lld seconds\n"
6047                            "average_speed_phase1: %llu items/sec\n"
6048                            "average_speed_phase2: N/A\n"
6049                            "real-time_speed_phase1: %llu items/sec\n"
6050                            "real-time_speed_phase2: N/A\n",
6051                            checked,
6052                            lo->ll_objs_checked_phase2,
6053                            rtime,
6054                            lo->ll_run_time_phase2,
6055                            speed,
6056                            new_checked);
6057
6058                 if (likely(lfsck->li_di_oit)) {
6059                         const struct dt_it_ops *iops =
6060                                 &lfsck->li_obj_oit->do_index_ops->dio_it;
6061
6062                         /* The low layer otable-based iteration position may NOT
6063                          * exactly match the layout-based directory traversal
6064                          * cookie. Generally, it is not a serious issue. But the
6065                          * caller should NOT make assumption on that. */
6066                         pos = iops->store(env, lfsck->li_di_oit);
6067                         if (!lfsck->li_current_oit_processed)
6068                                 pos--;
6069                 } else {
6070                         pos = lo->ll_pos_last_checkpoint;
6071                 }
6072
6073                 seq_printf(m, "current_position: %llu\n", pos);
6074         } else if (lo->ll_status == LS_SCANNING_PHASE2) {
6075                 time64_t duration = ktime_get_seconds() -
6076                                     com->lc_time_last_checkpoint;
6077                 u64 checked = lo->ll_objs_checked_phase2 +
6078                               com->lc_new_checked;
6079                 u64 speed1 = lo->ll_objs_checked_phase1;
6080                 u64 speed2 = checked;
6081                 u64 new_checked = com->lc_new_checked;
6082                 time64_t rtime = lo->ll_run_time_phase2 + duration;
6083
6084                 if (duration != 0)
6085                         new_checked = div64_s64(new_checked, duration);
6086                 if (lo->ll_run_time_phase1 != 0)
6087                         speed1 = div64_s64(speed1, lo->ll_run_time_phase1);
6088                 if (rtime != 0)
6089                         speed2 = div64_s64(speed2, rtime);
6090                 seq_printf(m, "checked_phase1: %llu\n"
6091                            "checked_phase2: %llu\n"
6092                            "run_time_phase1: %lld seconds\n"
6093                            "run_time_phase2: %lld seconds\n"
6094                            "average_speed_phase1: %llu items/sec\n"
6095                            "average_speed_phase2: %llu items/sec\n"
6096                            "real-time_speed_phase1: N/A\n"
6097                            "real-time_speed_phase2: %llu items/sec\n"
6098                            "current_position: "DFID"\n",
6099                            lo->ll_objs_checked_phase1,
6100                            checked,
6101                            lo->ll_run_time_phase1,
6102                            rtime,
6103                            speed1,
6104                            speed2,
6105                            new_checked,
6106                            PFID(&com->lc_fid_latest_scanned_phase2));
6107         } else {
6108                 __u64 speed1 = lo->ll_objs_checked_phase1;
6109                 __u64 speed2 = lo->ll_objs_checked_phase2;
6110
6111                 if (lo->ll_run_time_phase1 != 0)
6112                         speed1 = div64_s64(speed1, lo->ll_run_time_phase1);
6113                 if (lo->ll_run_time_phase2 != 0)
6114                         speed2 = div64_s64(speed2, lo->ll_run_time_phase2);
6115                 seq_printf(m, "checked_phase1: %llu\n"
6116                            "checked_phase2: %llu\n"
6117                            "run_time_phase1: %lld seconds\n"
6118                            "run_time_phase2: %lld seconds\n"
6119                            "average_speed_phase1: %llu items/sec\n"
6120                            "average_speed_phase2: %llu objs/sec\n"
6121                            "real-time_speed_phase1: N/A\n"
6122                            "real-time_speed_phase2: N/A\n"
6123                            "current_position: N/A\n",
6124                            lo->ll_objs_checked_phase1,
6125                            lo->ll_objs_checked_phase2,
6126                            lo->ll_run_time_phase1,
6127                            lo->ll_run_time_phase2,
6128                            speed1,
6129                            speed2);
6130         }
6131
6132         up_read(&com->lc_sem);
6133 }
6134
6135 static int lfsck_layout_master_double_scan(const struct lu_env *env,
6136                                            struct lfsck_component *com)
6137 {
6138         struct lfsck_layout             *lo     = com->lc_file_ram;
6139         struct lfsck_assistant_data     *lad    = com->lc_data;
6140         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6141         struct lfsck_tgt_descs          *ltds;
6142         struct lfsck_tgt_desc           *ltd;
6143         struct lfsck_tgt_desc           *next;
6144         int                              rc;
6145
6146         rc = lfsck_double_scan_generic(env, com, lo->ll_status);
6147
6148         if (thread_is_stopped(&lad->lad_thread)) {
6149                 LASSERT(list_empty(&lad->lad_req_list));
6150                 LASSERT(list_empty(&lad->lad_ost_phase1_list));
6151                 LASSERT(list_empty(&lad->lad_mdt_phase1_list));
6152
6153                 ltds = &lfsck->li_ost_descs;
6154                 spin_lock(&ltds->ltd_lock);
6155                 list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6156                                          ltd_layout_phase_list) {
6157                         list_del_init(&ltd->ltd_layout_phase_list);
6158                 }
6159                 spin_unlock(&ltds->ltd_lock);
6160
6161                 ltds = &lfsck->li_mdt_descs;
6162                 spin_lock(&ltds->ltd_lock);
6163                 list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6164                                          ltd_layout_phase_list) {
6165                         list_del_init(&ltd->ltd_layout_phase_list);
6166                 }
6167                 spin_unlock(&ltds->ltd_lock);
6168         }
6169
6170         return rc;
6171 }
6172
6173 static int lfsck_layout_slave_double_scan(const struct lu_env *env,
6174                                           struct lfsck_component *com)
6175 {
6176         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6177         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
6178         struct lfsck_layout             *lo     = com->lc_file_ram;
6179         struct ptlrpc_thread            *thread = &lfsck->li_thread;
6180         int                              rc;
6181         ENTRY;
6182
6183         CDEBUG(D_LFSCK, "%s: layout LFSCK slave phase2 scan start\n",
6184                lfsck_lfsck2name(lfsck));
6185
6186         atomic_inc(&lfsck->li_double_scan_count);
6187
6188         if (lo->ll_flags & LF_INCOMPLETE)
6189                 GOTO(done, rc = 1);
6190
6191         com->lc_new_checked = 0;
6192         com->lc_new_scanned = 0;
6193         com->lc_time_last_checkpoint = ktime_get_seconds();
6194         com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
6195                                        LFSCK_CHECKPOINT_INTERVAL;
6196
6197         while (1) {
6198                 rc = lfsck_layout_slave_query_master(env, com);
6199                 if (list_empty(&llsd->llsd_master_list)) {
6200                         if (unlikely(!thread_is_running(thread)))
6201                                 rc = 0;
6202                         else
6203                                 rc = 1;
6204
6205                         GOTO(done, rc);
6206                 }
6207
6208                 if (rc < 0)
6209                         GOTO(done, rc);
6210
6211                 rc = wait_event_idle_timeout(
6212                         thread->t_ctl_waitq,
6213                         !thread_is_running(thread) ||
6214                         lo->ll_flags & LF_INCOMPLETE ||
6215                         list_empty(&llsd->llsd_master_list),
6216                         cfs_time_seconds(30));
6217                 if (unlikely(!thread_is_running(thread)))
6218                         GOTO(done, rc = 0);
6219
6220                 if (lo->ll_flags & LF_INCOMPLETE)
6221                         GOTO(done, rc = 1);
6222
6223                 if (rc == 0)
6224                         continue;
6225
6226                 GOTO(done, rc = 1);
6227         }
6228
6229 done:
6230         rc = lfsck_layout_double_scan_result(env, com, rc);
6231         lfsck_layout_slave_notify_master(env, com, LE_PHASE2_DONE,
6232                         (rc > 0 && lo->ll_flags & LF_INCOMPLETE) ? 0 : rc);
6233         lfsck_layout_slave_quit(env, com);
6234         if (atomic_dec_and_test(&lfsck->li_double_scan_count))
6235                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
6236
6237         CDEBUG(D_LFSCK, "%s: layout LFSCK slave phase2 scan finished, "
6238                "status %d: rc = %d\n",
6239                lfsck_lfsck2name(lfsck), lo->ll_status, rc);
6240
6241         return rc;
6242 }
6243
6244 static void lfsck_layout_master_data_release(const struct lu_env *env,
6245                                              struct lfsck_component *com)
6246 {
6247         struct lfsck_assistant_data     *lad    = com->lc_data;
6248         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6249         struct lfsck_tgt_descs          *ltds;
6250         struct lfsck_tgt_desc           *ltd;
6251         struct lfsck_tgt_desc           *next;
6252
6253         LASSERT(lad != NULL);
6254         LASSERT(thread_is_init(&lad->lad_thread) ||
6255                 thread_is_stopped(&lad->lad_thread));
6256         LASSERT(list_empty(&lad->lad_req_list));
6257
6258         com->lc_data = NULL;
6259
6260         ltds = &lfsck->li_ost_descs;
6261         spin_lock(&ltds->ltd_lock);
6262         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase1_list,
6263                                  ltd_layout_phase_list) {
6264                 list_del_init(&ltd->ltd_layout_phase_list);
6265         }
6266         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6267                                  ltd_layout_phase_list) {
6268                 list_del_init(&ltd->ltd_layout_phase_list);
6269         }
6270         list_for_each_entry_safe(ltd, next, &lad->lad_ost_list,
6271                                  ltd_layout_list) {
6272                 list_del_init(&ltd->ltd_layout_list);
6273         }
6274         spin_unlock(&ltds->ltd_lock);
6275
6276         ltds = &lfsck->li_mdt_descs;
6277         spin_lock(&ltds->ltd_lock);
6278         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase1_list,
6279                                  ltd_layout_phase_list) {
6280                 list_del_init(&ltd->ltd_layout_phase_list);
6281         }
6282         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6283                                  ltd_layout_phase_list) {
6284                 list_del_init(&ltd->ltd_layout_phase_list);
6285         }
6286         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_list,
6287                                  ltd_layout_list) {
6288                 list_del_init(&ltd->ltd_layout_list);
6289         }
6290         spin_unlock(&ltds->ltd_lock);
6291
6292         if (likely(lad->lad_bitmap != NULL))
6293                 CFS_FREE_BITMAP(lad->lad_bitmap);
6294
6295         OBD_FREE_PTR(lad);
6296 }
6297
6298 static void lfsck_layout_slave_data_release(const struct lu_env *env,
6299                                             struct lfsck_component *com)
6300 {
6301         struct lfsck_layout_slave_data *llsd = com->lc_data;
6302
6303         lfsck_layout_slave_quit(env, com);
6304         com->lc_data = NULL;
6305         OBD_FREE_PTR(llsd);
6306 }
6307
6308 static void lfsck_layout_master_quit(const struct lu_env *env,
6309                                      struct lfsck_component *com)
6310 {
6311         struct lfsck_assistant_data     *lad    = com->lc_data;
6312         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6313         struct lfsck_tgt_descs          *ltds;
6314         struct lfsck_tgt_desc           *ltd;
6315         struct lfsck_tgt_desc           *next;
6316
6317         LASSERT(lad != NULL);
6318
6319         lfsck_quit_generic(env, com);
6320
6321         LASSERT(thread_is_init(&lad->lad_thread) ||
6322                 thread_is_stopped(&lad->lad_thread));
6323         LASSERT(list_empty(&lad->lad_req_list));
6324
6325         ltds = &lfsck->li_ost_descs;
6326         spin_lock(&ltds->ltd_lock);
6327         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase1_list,
6328                                  ltd_layout_phase_list) {
6329                 list_del_init(&ltd->ltd_layout_phase_list);
6330         }
6331         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6332                                  ltd_layout_phase_list) {
6333                 list_del_init(&ltd->ltd_layout_phase_list);
6334         }
6335         spin_unlock(&ltds->ltd_lock);
6336
6337         ltds = &lfsck->li_mdt_descs;
6338         spin_lock(&ltds->ltd_lock);
6339         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase1_list,
6340                                  ltd_layout_phase_list) {
6341                 list_del_init(&ltd->ltd_layout_phase_list);
6342         }
6343         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6344                                  ltd_layout_phase_list) {
6345                 list_del_init(&ltd->ltd_layout_phase_list);
6346         }
6347         spin_unlock(&ltds->ltd_lock);
6348 }
6349
6350 static void lfsck_layout_slave_quit(const struct lu_env *env,
6351                                     struct lfsck_component *com)
6352 {
6353         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
6354         struct lfsck_layout_seq          *lls;
6355         struct lfsck_layout_seq          *next;
6356         struct lfsck_layout_slave_target *llst;
6357
6358         LASSERT(llsd != NULL);
6359
6360         down_write(&com->lc_sem);
6361         list_for_each_entry_safe(lls, next, &llsd->llsd_seq_list,
6362                                  lls_list) {
6363                 list_del_init(&lls->lls_list);
6364                 lfsck_object_put(env, lls->lls_lastid_obj);
6365                 OBD_FREE_PTR(lls);
6366         }
6367         up_write(&com->lc_sem);
6368
6369         spin_lock(&llsd->llsd_lock);
6370         while (!list_empty(&llsd->llsd_master_list)) {
6371                 llst = list_entry(llsd->llsd_master_list.next,
6372                                   struct lfsck_layout_slave_target, llst_list);
6373                 list_del_init(&llst->llst_list);
6374                 spin_unlock(&llsd->llsd_lock);
6375                 lfsck_layout_llst_put(llst);
6376                 spin_lock(&llsd->llsd_lock);
6377         }
6378         spin_unlock(&llsd->llsd_lock);
6379
6380         lfsck_rbtree_cleanup(env, com);
6381 }
6382
6383 static int lfsck_layout_master_in_notify(const struct lu_env *env,
6384                                          struct lfsck_component *com,
6385                                          struct lfsck_request *lr)
6386 {
6387         struct lfsck_instance           *lfsck = com->lc_lfsck;
6388         struct lfsck_layout             *lo    = com->lc_file_ram;
6389         struct lfsck_assistant_data     *lad   = com->lc_data;
6390         struct lfsck_tgt_descs          *ltds;
6391         struct lfsck_tgt_desc           *ltd;
6392         bool                             fail  = false;
6393         ENTRY;
6394
6395         if (lr->lr_event == LE_PAIRS_VERIFY) {
6396                 int rc;
6397
6398                 rc = lfsck_layout_master_check_pairs(env, com, &lr->lr_fid,
6399                                                      &lr->lr_fid2,
6400                                                      lr->lr_comp_id);
6401
6402                 RETURN(rc);
6403         }
6404
6405         CDEBUG(D_LFSCK, "%s: layout LFSCK master handles notify %u "
6406                "from %s %x, status %d, flags %x, flags2 %x\n",
6407                lfsck_lfsck2name(lfsck), lr->lr_event,
6408                (lr->lr_flags & LEF_FROM_OST) ? "OST" : "MDT",
6409                lr->lr_index, lr->lr_status, lr->lr_flags, lr->lr_flags2);
6410
6411         if (lr->lr_event != LE_PHASE1_DONE &&
6412             lr->lr_event != LE_PHASE2_DONE &&
6413             lr->lr_event != LE_PEER_EXIT)
6414                 RETURN(-EINVAL);
6415
6416         if (lr->lr_flags & LEF_FROM_OST)
6417                 ltds = &lfsck->li_ost_descs;
6418         else
6419                 ltds = &lfsck->li_mdt_descs;
6420         spin_lock(&ltds->ltd_lock);
6421         ltd = lfsck_ltd2tgt(ltds, lr->lr_index);
6422         if (ltd == NULL) {
6423                 spin_unlock(&ltds->ltd_lock);
6424
6425                 RETURN(-ENXIO);
6426         }
6427
6428         list_del_init(&ltd->ltd_layout_phase_list);
6429         switch (lr->lr_event) {
6430         case LE_PHASE1_DONE:
6431                 if (lr->lr_status <= 0 || lr->lr_flags2 & LF_INCOMPLETE) {
6432                         if (lr->lr_flags2 & LF_INCOMPLETE) {
6433                                 if (lr->lr_flags & LEF_FROM_OST)
6434                                         lfsck_lad_set_bitmap(env, com,
6435                                                              ltd->ltd_index);
6436                                 else
6437                                         lo->ll_flags |= LF_INCOMPLETE;
6438                         }
6439                         ltd->ltd_layout_done = 1;
6440                         list_del_init(&ltd->ltd_layout_list);
6441                         fail = true;
6442                         break;
6443                 }
6444
6445                 if (lr->lr_flags & LEF_FROM_OST) {
6446                         if (list_empty(&ltd->ltd_layout_list))
6447                                 list_add_tail(&ltd->ltd_layout_list,
6448                                               &lad->lad_ost_list);
6449                         list_add_tail(&ltd->ltd_layout_phase_list,
6450                                       &lad->lad_ost_phase2_list);
6451                 } else {
6452                         if (list_empty(&ltd->ltd_layout_list))
6453                                 list_add_tail(&ltd->ltd_layout_list,
6454                                               &lad->lad_mdt_list);
6455                         list_add_tail(&ltd->ltd_layout_phase_list,
6456                                       &lad->lad_mdt_phase2_list);
6457                 }
6458                 break;
6459         case LE_PHASE2_DONE:
6460                 ltd->ltd_layout_done = 1;
6461                 if (!list_empty(&ltd->ltd_layout_list))
6462                         list_del_init(&ltd->ltd_layout_list);
6463
6464                 if (lr->lr_flags2 & LF_INCOMPLETE) {
6465                         lfsck_lad_set_bitmap(env, com, ltd->ltd_index);
6466                         fail = true;
6467                 }
6468
6469                 break;
6470         case LE_PEER_EXIT:
6471                 fail = true;
6472                 ltd->ltd_layout_done = 1;
6473                 list_del_init(&ltd->ltd_layout_list);
6474                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT) &&
6475                     !(lr->lr_flags & LEF_FROM_OST))
6476                                 lo->ll_flags |= LF_INCOMPLETE;
6477                 break;
6478         default:
6479                 break;
6480         }
6481         spin_unlock(&ltds->ltd_lock);
6482
6483         if (fail && lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT) {
6484                 struct lfsck_stop *stop = &lfsck_env_info(env)->lti_stop;
6485
6486                 memset(stop, 0, sizeof(*stop));
6487                 stop->ls_status = lr->lr_status;
6488                 stop->ls_flags = lr->lr_param & ~LPF_BROADCAST;
6489                 lfsck_stop(env, lfsck->li_bottom, stop);
6490         } else if (lfsck_phase2_next_ready(lad)) {
6491                 wake_up_all(&lad->lad_thread.t_ctl_waitq);
6492         }
6493
6494         RETURN(0);
6495 }
6496
6497 static int lfsck_layout_slave_in_notify_local(const struct lu_env *env,
6498                                               struct lfsck_component *com,
6499                                               struct lfsck_req_local *lrl,
6500                                               struct thandle *th)
6501 {
6502         ENTRY;
6503
6504         switch (lrl->lrl_event) {
6505         case LEL_FID_ACCESSED:
6506                 lfsck_rbtree_update_bitmap(env, com, &lrl->lrl_fid, true);
6507                 RETURN(0);
6508         case LEL_PAIRS_VERIFY_LOCAL: {
6509                 int rc;
6510
6511                 lrl->lrl_status = LPVS_INIT;
6512                 /* Firstly, if the MDT-object which is claimed via OST-object
6513                  * local stored PFID xattr recognizes the OST-object, then it
6514                  * must be that the client given PFID is wrong. */
6515                 rc = lfsck_layout_slave_check_pairs(env, com, &lrl->lrl_fid,
6516                                 &lrl->lrl_ff_local.ff_parent,
6517                                 lrl->lrl_ff_local.ff_layout.ol_comp_id);
6518                 if (rc <= 0)
6519                         RETURN(0);
6520
6521                 lrl->lrl_status = LPVS_INCONSISTENT;
6522                 /* The OST-object local stored PFID xattr is stale. We need to
6523                  * check whether the MDT-object that is claimed via the client
6524                  * given PFID information recognizes the OST-object or not. If
6525                  * matches, then need to update the OST-object's PFID xattr. */
6526                 rc = lfsck_layout_slave_check_pairs(env, com, &lrl->lrl_fid,
6527                                 &lrl->lrl_ff_client.ff_parent,
6528                                 lrl->lrl_ff_client.ff_layout.ol_comp_id);
6529                 /* For rc < 0 case:
6530                  * We are not sure whether the client given PFID information
6531                  * is correct or not, do nothing to avoid improper fixing.
6532                  *
6533                  * For rc > 0 case:
6534                  * The client given PFID information is also invalid, we can
6535                  * NOT fix the OST-object inconsistency.
6536                  */
6537                 if (!rc) {
6538                         lrl->lrl_status = LPVS_INCONSISTENT_TOFIX;
6539                         rc = lfsck_layout_slave_repair_pfid(env, com, lrl);
6540                 }
6541
6542                 RETURN(rc);
6543         }
6544         default:
6545                 break;
6546         }
6547
6548         RETURN(-EOPNOTSUPP);
6549 }
6550
6551 static int lfsck_layout_slave_in_notify(const struct lu_env *env,
6552                                         struct lfsck_component *com,
6553                                         struct lfsck_request *lr)
6554 {
6555         struct lfsck_instance *lfsck = com->lc_lfsck;
6556         struct lfsck_layout_slave_data *llsd = com->lc_data;
6557         struct lfsck_layout_slave_target *llst;
6558         int rc;
6559         ENTRY;
6560
6561         switch (lr->lr_event) {
6562         case LE_CONDITIONAL_DESTROY:
6563                 rc = lfsck_layout_slave_conditional_destroy(env, com, lr);
6564                 RETURN(rc);
6565         case LE_PHASE1_DONE: {
6566                 if (lr->lr_flags2 & LF_INCOMPLETE) {
6567                         struct lfsck_layout *lo = com->lc_file_ram;
6568
6569                         lo->ll_flags |= LF_INCOMPLETE;
6570                         llst = lfsck_layout_llst_find_and_del(llsd,
6571                                                               lr->lr_index,
6572                                                               true);
6573                         if (llst != NULL) {
6574                                 lfsck_layout_llst_put(llst);
6575                                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
6576                         }
6577                 }
6578
6579                 RETURN(0);
6580         }
6581         case LE_PHASE2_DONE:
6582         case LE_PEER_EXIT:
6583                 CDEBUG(D_LFSCK, "%s: layout LFSCK slave handle notify %u "
6584                        "from MDT %x, status %d\n", lfsck_lfsck2name(lfsck),
6585                        lr->lr_event, lr->lr_index, lr->lr_status);
6586                 break;
6587         default:
6588                 RETURN(-EINVAL);
6589         }
6590
6591         llst = lfsck_layout_llst_find_and_del(llsd, lr->lr_index, true);
6592         if (llst == NULL)
6593                 RETURN(0);
6594
6595         lfsck_layout_llst_put(llst);
6596         if (list_empty(&llsd->llsd_master_list))
6597                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
6598
6599         if (lr->lr_event == LE_PEER_EXIT &&
6600             (lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT ||
6601              (list_empty(&llsd->llsd_master_list) &&
6602               (lr->lr_status == LS_STOPPED ||
6603                lr->lr_status == LS_CO_STOPPED)))) {
6604                 struct lfsck_stop *stop = &lfsck_env_info(env)->lti_stop;
6605
6606                 memset(stop, 0, sizeof(*stop));
6607                 stop->ls_status = lr->lr_status;
6608                 stop->ls_flags = lr->lr_param & ~LPF_BROADCAST;
6609                 lfsck_stop(env, lfsck->li_bottom, stop);
6610         }
6611
6612         RETURN(0);
6613 }
6614
6615 static void lfsck_layout_repaired(struct lfsck_layout *lo, __u64 *count)
6616 {
6617         int i;
6618
6619         for (i = 0; i < LLIT_MAX; i++)
6620                 *count += lo->ll_objs_repaired[i];
6621 }
6622
6623 static int lfsck_layout_query_all(const struct lu_env *env,
6624                                   struct lfsck_component *com,
6625                                   __u32 *mdts_count, __u32 *osts_count,
6626                                   __u64 *repaired)
6627 {
6628         struct lfsck_layout *lo = com->lc_file_ram;
6629         struct lfsck_tgt_descs *ltds;
6630         struct lfsck_tgt_desc *ltd;
6631         int idx;
6632         int rc;
6633         ENTRY;
6634
6635         rc = lfsck_query_all(env, com);
6636         if (rc != 0)
6637                 RETURN(rc);
6638
6639         ltds = &com->lc_lfsck->li_mdt_descs;
6640         down_read(&ltds->ltd_rw_sem);
6641         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
6642                 ltd = lfsck_ltd2tgt(ltds, idx);
6643                 LASSERT(ltd != NULL);
6644
6645                 mdts_count[ltd->ltd_layout_status]++;
6646                 *repaired += ltd->ltd_layout_repaired;
6647         }
6648         up_read(&ltds->ltd_rw_sem);
6649
6650         ltds = &com->lc_lfsck->li_ost_descs;
6651         down_read(&ltds->ltd_rw_sem);
6652         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
6653                 ltd = lfsck_ltd2tgt(ltds, idx);
6654                 LASSERT(ltd != NULL);
6655
6656                 osts_count[ltd->ltd_layout_status]++;
6657                 *repaired += ltd->ltd_layout_repaired;
6658         }
6659         up_read(&ltds->ltd_rw_sem);
6660
6661         down_read(&com->lc_sem);
6662         mdts_count[lo->ll_status]++;
6663         lfsck_layout_repaired(lo, repaired);
6664         up_read(&com->lc_sem);
6665
6666         RETURN(0);
6667 }
6668
6669 static int lfsck_layout_query(const struct lu_env *env,
6670                               struct lfsck_component *com,
6671                               struct lfsck_request *req,
6672                               struct lfsck_reply *rep,
6673                               struct lfsck_query *que, int idx)
6674 {
6675         struct lfsck_layout *lo = com->lc_file_ram;
6676         int rc = 0;
6677
6678         if (que != NULL) {
6679                 LASSERT(com->lc_lfsck->li_master);
6680
6681                 rc = lfsck_layout_query_all(env, com,
6682                                             que->lu_mdts_count[idx],
6683                                             que->lu_osts_count[idx],
6684                                             &que->lu_repaired[idx]);
6685         } else {
6686                 down_read(&com->lc_sem);
6687                 rep->lr_status = lo->ll_status;
6688                 if (req->lr_flags & LEF_QUERY_ALL)
6689                         lfsck_layout_repaired(lo, &rep->lr_repaired);
6690                 up_read(&com->lc_sem);
6691         }
6692
6693         return rc;
6694 }
6695
6696 /* with lfsck::li_lock held */
6697 static int lfsck_layout_slave_join(const struct lu_env *env,
6698                                    struct lfsck_component *com,
6699                                    struct lfsck_start_param *lsp)
6700 {
6701         struct lfsck_instance            *lfsck = com->lc_lfsck;
6702         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
6703         struct lfsck_layout_slave_target *llst;
6704         struct lfsck_start               *start = lsp->lsp_start;
6705         int                               rc    = 0;
6706         ENTRY;
6707
6708         if (start == NULL || !(start->ls_flags & LPF_OST_ORPHAN))
6709                 RETURN(0);
6710
6711         if (!lsp->lsp_index_valid)
6712                 RETURN(-EINVAL);
6713
6714         /* If someone is running the LFSCK without orphan handling,
6715          * it will not maintain the object accessing rbtree. So we
6716          * cannot join it for orphan handling. */
6717         if (!llsd->llsd_rbtree_valid)
6718                 RETURN(-EBUSY);
6719
6720         spin_unlock(&lfsck->li_lock);
6721         rc = lfsck_layout_llst_add(llsd, lsp->lsp_index);
6722         spin_lock(&lfsck->li_lock);
6723         if (rc == 0 && !thread_is_running(&lfsck->li_thread)) {
6724                 spin_unlock(&lfsck->li_lock);
6725                 llst = lfsck_layout_llst_find_and_del(llsd, lsp->lsp_index,
6726                                                       true);
6727                 if (llst != NULL)
6728                         lfsck_layout_llst_put(llst);
6729                 spin_lock(&lfsck->li_lock);
6730                 rc = -EAGAIN;
6731         }
6732
6733         RETURN(rc);
6734 }
6735
6736 static struct lfsck_operations lfsck_layout_master_ops = {
6737         .lfsck_reset            = lfsck_layout_reset,
6738         .lfsck_fail             = lfsck_layout_fail,
6739         .lfsck_checkpoint       = lfsck_layout_master_checkpoint,
6740         .lfsck_prep             = lfsck_layout_master_prep,
6741         .lfsck_exec_oit         = lfsck_layout_master_exec_oit,
6742         .lfsck_exec_dir         = lfsck_layout_exec_dir,
6743         .lfsck_post             = lfsck_layout_master_post,
6744         .lfsck_dump             = lfsck_layout_dump,
6745         .lfsck_double_scan      = lfsck_layout_master_double_scan,
6746         .lfsck_data_release     = lfsck_layout_master_data_release,
6747         .lfsck_quit             = lfsck_layout_master_quit,
6748         .lfsck_in_notify        = lfsck_layout_master_in_notify,
6749         .lfsck_query            = lfsck_layout_query,
6750 };
6751
6752 static struct lfsck_operations lfsck_layout_slave_ops = {
6753         .lfsck_reset            = lfsck_layout_reset,
6754         .lfsck_fail             = lfsck_layout_fail,
6755         .lfsck_checkpoint       = lfsck_layout_slave_checkpoint,
6756         .lfsck_prep             = lfsck_layout_slave_prep,
6757         .lfsck_exec_oit         = lfsck_layout_slave_exec_oit,
6758         .lfsck_exec_dir         = lfsck_layout_exec_dir,
6759         .lfsck_post             = lfsck_layout_slave_post,
6760         .lfsck_dump             = lfsck_layout_dump,
6761         .lfsck_double_scan      = lfsck_layout_slave_double_scan,
6762         .lfsck_data_release     = lfsck_layout_slave_data_release,
6763         .lfsck_quit             = lfsck_layout_slave_quit,
6764         .lfsck_in_notify_local  = lfsck_layout_slave_in_notify_local,
6765         .lfsck_in_notify        = lfsck_layout_slave_in_notify,
6766         .lfsck_query            = lfsck_layout_query,
6767         .lfsck_join             = lfsck_layout_slave_join,
6768 };
6769
6770 static void lfsck_layout_assistant_fill_pos(const struct lu_env *env,
6771                                             struct lfsck_component *com,
6772                                             struct lfsck_position *pos)
6773 {
6774         struct lfsck_assistant_data     *lad = com->lc_data;
6775         struct lfsck_layout_req         *llr;
6776
6777         if (((struct lfsck_layout *)(com->lc_file_ram))->ll_status !=
6778             LS_SCANNING_PHASE1)
6779                 return;
6780
6781         if (list_empty(&lad->lad_req_list))
6782                 return;
6783
6784         llr = list_entry(lad->lad_req_list.next,
6785                          struct lfsck_layout_req,
6786                          llr_lar.lar_list);
6787         pos->lp_oit_cookie = llr->llr_lar.lar_parent->lso_oit_cookie - 1;
6788 }
6789
6790 struct lfsck_assistant_operations lfsck_layout_assistant_ops = {
6791         .la_handler_p1          = lfsck_layout_assistant_handler_p1,
6792         .la_handler_p2          = lfsck_layout_assistant_handler_p2,
6793         .la_fill_pos            = lfsck_layout_assistant_fill_pos,
6794         .la_double_scan_result  = lfsck_layout_double_scan_result,
6795         .la_req_fini            = lfsck_layout_assistant_req_fini,
6796         .la_sync_failures       = lfsck_layout_assistant_sync_failures,
6797 };
6798
6799 int lfsck_layout_setup(const struct lu_env *env, struct lfsck_instance *lfsck)
6800 {
6801         struct lfsck_component  *com;
6802         struct lfsck_layout     *lo;
6803         struct dt_object        *root = NULL;
6804         struct dt_object        *obj;
6805         int                      i;
6806         int                      rc;
6807         ENTRY;
6808
6809         OBD_ALLOC_PTR(com);
6810         if (com == NULL)
6811                 RETURN(-ENOMEM);
6812
6813         INIT_LIST_HEAD(&com->lc_link);
6814         INIT_LIST_HEAD(&com->lc_link_dir);
6815         init_rwsem(&com->lc_sem);
6816         atomic_set(&com->lc_ref, 1);
6817         com->lc_lfsck = lfsck;
6818         com->lc_type = LFSCK_TYPE_LAYOUT;
6819         if (lfsck->li_master) {
6820                 com->lc_ops = &lfsck_layout_master_ops;
6821                 com->lc_data = lfsck_assistant_data_init(
6822                                 &lfsck_layout_assistant_ops,
6823                                 LFSCK_LAYOUT);
6824                 if (com->lc_data == NULL)
6825                         GOTO(out, rc = -ENOMEM);
6826
6827                 for (i = 0; i < LFSCK_STF_COUNT; i++)
6828                         mutex_init(&com->lc_sub_trace_objs[i].lsto_mutex);
6829         } else {
6830                 struct lfsck_layout_slave_data *llsd;
6831
6832                 com->lc_ops = &lfsck_layout_slave_ops;
6833                 OBD_ALLOC_PTR(llsd);
6834                 if (llsd == NULL)
6835                         GOTO(out, rc = -ENOMEM);
6836
6837                 INIT_LIST_HEAD(&llsd->llsd_seq_list);
6838                 INIT_LIST_HEAD(&llsd->llsd_master_list);
6839                 spin_lock_init(&llsd->llsd_lock);
6840                 llsd->llsd_rb_root = RB_ROOT;
6841                 init_rwsem(&llsd->llsd_rb_rwsem);
6842                 com->lc_data = llsd;
6843         }
6844         com->lc_file_size = sizeof(*lo);
6845         OBD_ALLOC(com->lc_file_ram, com->lc_file_size);
6846         if (com->lc_file_ram == NULL)
6847                 GOTO(out, rc = -ENOMEM);
6848
6849         OBD_ALLOC(com->lc_file_disk, com->lc_file_size);
6850         if (com->lc_file_disk == NULL)
6851                 GOTO(out, rc = -ENOMEM);
6852
6853         root = dt_locate(env, lfsck->li_bottom, &lfsck->li_local_root_fid);
6854         if (IS_ERR(root))
6855                 GOTO(out, rc = PTR_ERR(root));
6856
6857         if (unlikely(!dt_try_as_dir(env, root)))
6858                 GOTO(out, rc = -ENOTDIR);
6859
6860         obj = local_file_find_or_create(env, lfsck->li_los, root,
6861                                         LFSCK_LAYOUT,
6862                                         S_IFREG | S_IRUGO | S_IWUSR);
6863         if (IS_ERR(obj))
6864                 GOTO(out, rc = PTR_ERR(obj));
6865
6866         com->lc_obj = obj;
6867         rc = lfsck_layout_load(env, com);
6868         if (rc > 0) {
6869                 rc = lfsck_layout_reset(env, com, true);
6870         } else if (rc == -ENOENT) {
6871                 rc = lfsck_layout_init(env, com);
6872         } else if (lfsck->li_master) {
6873                 rc = lfsck_load_sub_trace_files(env, com,
6874                                 &dt_lfsck_layout_dangling_features,
6875                                 LFSCK_LAYOUT, false);
6876                 if (rc)
6877                         rc = lfsck_layout_reset(env, com, true);
6878         }
6879
6880         if (rc != 0)
6881                 GOTO(out, rc);
6882
6883         lo = com->lc_file_ram;
6884         switch (lo->ll_status) {
6885         case LS_INIT:
6886         case LS_COMPLETED:
6887         case LS_FAILED:
6888         case LS_STOPPED:
6889         case LS_PARTIAL:
6890                 spin_lock(&lfsck->li_lock);
6891                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
6892                 spin_unlock(&lfsck->li_lock);
6893                 break;
6894         default:
6895                 CERROR("%s: unknown lfsck_layout status %d\n",
6896                        lfsck_lfsck2name(lfsck), lo->ll_status);
6897                 /* fall through */
6898         case LS_SCANNING_PHASE1:
6899         case LS_SCANNING_PHASE2:
6900                 /* No need to store the status to disk right now.
6901                  * If the system crashed before the status stored,
6902                  * it will be loaded back when next time. */
6903                 lo->ll_status = LS_CRASHED;
6904                 if (!lfsck->li_master)
6905                         lo->ll_flags |= LF_INCOMPLETE;
6906                 /* fall through */
6907         case LS_PAUSED:
6908         case LS_CRASHED:
6909         case LS_CO_FAILED:
6910         case LS_CO_STOPPED:
6911         case LS_CO_PAUSED:
6912                 spin_lock(&lfsck->li_lock);
6913                 list_add_tail(&com->lc_link, &lfsck->li_list_scan);
6914                 spin_unlock(&lfsck->li_lock);
6915                 break;
6916         }
6917
6918         if (lo->ll_flags & LF_CRASHED_LASTID) {
6919                 LASSERT(lfsck->li_out_notify != NULL);
6920
6921                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
6922                                      LE_LASTID_REBUILDING);
6923         }
6924
6925         GOTO(out, rc = 0);
6926
6927 out:
6928         if (root != NULL && !IS_ERR(root))
6929                 lfsck_object_put(env, root);
6930
6931         if (rc != 0) {
6932                 lfsck_component_cleanup(env, com);
6933                 CERROR("%s: fail to init layout LFSCK component: rc = %d\n",
6934                        lfsck_lfsck2name(lfsck), rc);
6935         }
6936
6937         return rc;
6938 }
6939
6940 struct lfsck_orphan_it {
6941         struct lfsck_component           *loi_com;
6942         struct lfsck_rbtree_node         *loi_lrn;
6943         struct lfsck_layout_slave_target *loi_llst;
6944         struct lu_fid                     loi_key;
6945         struct lu_orphan_rec_v3           loi_rec;
6946         __u64                             loi_hash;
6947         unsigned int                      loi_over:1;
6948 };
6949
6950 static int lfsck_fid_match_idx(const struct lu_env *env,
6951                                struct lfsck_instance *lfsck,
6952                                const struct lu_fid *fid, int idx)
6953 {
6954         struct seq_server_site  *ss;
6955         struct lu_server_fld    *sf;
6956         struct lu_seq_range     *range = &lfsck_env_info(env)->lti_range;
6957         int                      rc;
6958
6959         /* All abnormal cases will be returned to MDT0. */
6960         if (!fid_is_norm(fid)) {
6961                 if (idx == 0)
6962                         return 1;
6963
6964                 return 0;
6965         }
6966
6967         ss = lfsck_dev_site(lfsck);
6968         if (unlikely(ss == NULL))
6969                 return -ENOTCONN;
6970
6971         sf = ss->ss_server_fld;
6972         LASSERT(sf != NULL);
6973
6974         fld_range_set_any(range);
6975         rc = fld_server_lookup(env, sf, fid_seq(fid), range);
6976         if (rc != 0)
6977                 return rc;
6978
6979         if (!fld_range_is_mdt(range))
6980                 return -EINVAL;
6981
6982         if (range->lsr_index == idx)
6983                 return 1;
6984
6985         return 0;
6986 }
6987
6988 static void lfsck_layout_destroy_orphan(const struct lu_env *env,
6989                                         struct dt_object *obj)
6990 {
6991         struct dt_device        *dev    = lfsck_obj2dev(obj);
6992         struct thandle          *handle;
6993         int                      rc;
6994         ENTRY;
6995
6996         handle = dt_trans_create(env, dev);
6997         if (IS_ERR(handle))
6998                 RETURN_EXIT;
6999
7000         rc = dt_declare_ref_del(env, obj, handle);
7001         if (rc != 0)
7002                 GOTO(stop, rc);
7003
7004         rc = dt_declare_destroy(env, obj, handle);
7005         if (rc != 0)
7006                 GOTO(stop, rc);
7007
7008         rc = dt_trans_start_local(env, dev, handle);
7009         if (rc != 0)
7010                 GOTO(stop, rc);
7011
7012         dt_write_lock(env, obj, 0);
7013         rc = dt_ref_del(env, obj, handle);
7014         if (rc == 0)
7015                 rc = dt_destroy(env, obj, handle);
7016         dt_write_unlock(env, obj);
7017
7018         GOTO(stop, rc);
7019
7020 stop:
7021         dt_trans_stop(env, dev, handle);
7022
7023         CDEBUG(D_LFSCK, "destroy orphan OST-object "DFID": rc = %d\n",
7024                PFID(lfsck_dto2fid(obj)), rc);
7025
7026         RETURN_EXIT;
7027 }
7028
7029 static int lfsck_orphan_index_lookup(const struct lu_env *env,
7030                                      struct dt_object *dt,
7031                                      struct dt_rec *rec,
7032                                      const struct dt_key *key)
7033 {
7034         return -EOPNOTSUPP;
7035 }
7036
7037 static int lfsck_orphan_index_declare_insert(const struct lu_env *env,
7038                                              struct dt_object *dt,
7039                                              const struct dt_rec *rec,
7040                                              const struct dt_key *key,
7041                                              struct thandle *handle)
7042 {
7043         return -EOPNOTSUPP;
7044 }
7045
7046 static int lfsck_orphan_index_insert(const struct lu_env *env,
7047                                      struct dt_object *dt,
7048                                      const struct dt_rec *rec,
7049                                      const struct dt_key *key,
7050                                      struct thandle *handle)
7051 {
7052         return -EOPNOTSUPP;
7053 }
7054
7055 static int lfsck_orphan_index_declare_delete(const struct lu_env *env,
7056                                              struct dt_object *dt,
7057                                              const struct dt_key *key,
7058                                              struct thandle *handle)
7059 {
7060         return -EOPNOTSUPP;
7061 }
7062
7063 static int lfsck_orphan_index_delete(const struct lu_env *env,
7064                                      struct dt_object *dt,
7065                                      const struct dt_key *key,
7066                                      struct thandle *handle)
7067 {
7068         return -EOPNOTSUPP;
7069 }
7070
7071 static struct dt_it *lfsck_orphan_it_init(const struct lu_env *env,
7072                                           struct dt_object *dt,
7073                                           __u32 attr)
7074 {
7075         struct dt_device                *dev    = lu2dt_dev(dt->do_lu.lo_dev);
7076         struct lfsck_instance           *lfsck;
7077         struct lfsck_component          *com    = NULL;
7078         struct lfsck_layout_slave_data  *llsd;
7079         struct lfsck_orphan_it          *it     = NULL;
7080         struct lfsck_layout             *lo;
7081         int                              rc     = 0;
7082         ENTRY;
7083
7084         lfsck = lfsck_instance_find(dev, true, false);
7085         if (unlikely(lfsck == NULL))
7086                 RETURN(ERR_PTR(-ENXIO));
7087
7088         com = lfsck_component_find(lfsck, LFSCK_TYPE_LAYOUT);
7089         if (unlikely(com == NULL))
7090                 GOTO(out, rc = -ENOENT);
7091
7092         lo = com->lc_file_ram;
7093         if (lo->ll_flags & LF_INCOMPLETE)
7094                 GOTO(out, rc = -ESRCH);
7095
7096         llsd = com->lc_data;
7097         if (!llsd->llsd_rbtree_valid)
7098                 GOTO(out, rc = -ESRCH);
7099
7100         OBD_ALLOC_PTR(it);
7101         if (it == NULL)
7102                 GOTO(out, rc = -ENOMEM);
7103
7104         it->loi_llst = lfsck_layout_llst_find_and_del(llsd, attr, false);
7105         if (it->loi_llst == NULL)
7106                 GOTO(out, rc = -ENXIO);
7107
7108         if (dev->dd_record_fid_accessed) {
7109                 /* The first iteration against the rbtree, scan the whole rbtree
7110                  * to remove the nodes which do NOT need to be handled. */
7111                 down_write(&llsd->llsd_rb_rwsem);
7112                 if (dev->dd_record_fid_accessed) {
7113                         struct rb_node                  *node;
7114                         struct rb_node                  *next;
7115                         struct lfsck_rbtree_node        *lrn;
7116
7117                         /* No need to record the fid accessing anymore. */
7118                         dev->dd_record_fid_accessed = 0;
7119
7120                         node = rb_first(&llsd->llsd_rb_root);
7121                         while (node != NULL) {
7122                                 next = rb_next(node);
7123                                 lrn = rb_entry(node, struct lfsck_rbtree_node,
7124                                                lrn_node);
7125                                 if (atomic_read(&lrn->lrn_known_count) <=
7126                                     atomic_read(&lrn->lrn_accessed_count)) {
7127                                         rb_erase(node, &llsd->llsd_rb_root);
7128                                         lfsck_rbtree_free(lrn);
7129                                 }
7130                                 node = next;
7131                         }
7132                 }
7133                 up_write(&llsd->llsd_rb_rwsem);
7134         }
7135
7136         /* read lock the rbtree when init, and unlock when fini */
7137         down_read(&llsd->llsd_rb_rwsem);
7138         it->loi_com = com;
7139         com = NULL;
7140
7141         GOTO(out, rc = 0);
7142
7143 out:
7144         if (com != NULL)
7145                 lfsck_component_put(env, com);
7146
7147         CDEBUG(D_LFSCK, "%s: init the orphan iteration: rc = %d\n",
7148                lfsck_lfsck2name(lfsck), rc);
7149
7150         lfsck_instance_put(env, lfsck);
7151         if (rc != 0) {
7152                 if (it != NULL)
7153                         OBD_FREE_PTR(it);
7154
7155                 it = (struct lfsck_orphan_it *)ERR_PTR(rc);
7156         }
7157
7158         return (struct dt_it *)it;
7159 }
7160
7161 static void lfsck_orphan_it_fini(const struct lu_env *env,
7162                                  struct dt_it *di)
7163 {
7164         struct lfsck_orphan_it           *it    = (struct lfsck_orphan_it *)di;
7165         struct lfsck_component           *com   = it->loi_com;
7166         struct lfsck_layout_slave_data   *llsd;
7167         struct lfsck_layout_slave_target *llst;
7168
7169         if (com != NULL) {
7170                 CDEBUG(D_LFSCK, "%s: fini the orphan iteration\n",
7171                        lfsck_lfsck2name(com->lc_lfsck));
7172
7173                 llsd = com->lc_data;
7174                 up_read(&llsd->llsd_rb_rwsem);
7175                 llst = it->loi_llst;
7176                 LASSERT(llst != NULL);
7177
7178                 /* Save the key and hash for iterate next. */
7179                 llst->llst_fid = it->loi_key;
7180                 llst->llst_hash = it->loi_hash;
7181                 lfsck_layout_llst_put(llst);
7182                 lfsck_component_put(env, com);
7183         }
7184         OBD_FREE_PTR(it);
7185 }
7186
7187 /**
7188  * \retval       +1: the iteration finished
7189  * \retval        0: on success, not finished
7190  * \retval      -ve: on error
7191  */
7192 static int lfsck_orphan_it_next(const struct lu_env *env,
7193                                 struct dt_it *di)
7194 {
7195         struct lfsck_thread_info        *info   = lfsck_env_info(env);
7196         struct filter_fid               *ff     = &info->lti_ff;
7197         struct lu_attr                  *la     = &info->lti_la;
7198         struct lfsck_orphan_it          *it     = (struct lfsck_orphan_it *)di;
7199         struct lu_fid                   *key    = &it->loi_key;
7200         struct lu_orphan_rec_v3         *rec    = &it->loi_rec;
7201         struct ost_layout               *ol     = &rec->lor_layout;
7202         struct lfsck_component          *com    = it->loi_com;
7203         struct lfsck_instance           *lfsck  = com->lc_lfsck;
7204         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
7205         struct dt_object                *obj;
7206         struct lfsck_rbtree_node        *lrn;
7207         int                              pos;
7208         int                              rc;
7209         __u32                            save;
7210         __u32                            idx    = it->loi_llst->llst_index;
7211         bool                             exact  = false;
7212         ENTRY;
7213
7214         if (it->loi_over)
7215                 RETURN(1);
7216
7217 again0:
7218         lrn = it->loi_lrn;
7219         if (lrn == NULL) {
7220                 lrn = lfsck_rbtree_search(llsd, key, &exact);
7221                 if (lrn == NULL) {
7222                         it->loi_over = 1;
7223                         RETURN(1);
7224                 }
7225
7226                 it->loi_lrn = lrn;
7227                 if (!exact) {
7228                         key->f_seq = lrn->lrn_seq;
7229                         key->f_oid = lrn->lrn_first_oid;
7230                         key->f_ver = 0;
7231                 }
7232         } else {
7233                 key->f_oid++;
7234                 if (unlikely(key->f_oid == 0)) {
7235                         key->f_seq++;
7236                         it->loi_lrn = NULL;
7237                         goto again0;
7238                 }
7239
7240                 if (key->f_oid >=
7241                     lrn->lrn_first_oid + LFSCK_RBTREE_BITMAP_WIDTH) {
7242                         it->loi_lrn = NULL;
7243                         goto again0;
7244                 }
7245         }
7246
7247         if (unlikely(atomic_read(&lrn->lrn_known_count) <=
7248                      atomic_read(&lrn->lrn_accessed_count))) {
7249                 struct rb_node *next = rb_next(&lrn->lrn_node);
7250
7251                 while (next != NULL) {
7252                         lrn = rb_entry(next, struct lfsck_rbtree_node,
7253                                        lrn_node);
7254                         if (atomic_read(&lrn->lrn_known_count) >
7255                             atomic_read(&lrn->lrn_accessed_count))
7256                                 break;
7257                         next = rb_next(next);
7258                 }
7259
7260                 if (next == NULL) {
7261                         it->loi_over = 1;
7262                         RETURN(1);
7263                 }
7264
7265                 it->loi_lrn = lrn;
7266                 key->f_seq = lrn->lrn_seq;
7267                 key->f_oid = lrn->lrn_first_oid;
7268                 key->f_ver = 0;
7269         }
7270
7271         pos = key->f_oid - lrn->lrn_first_oid;
7272
7273 again1:
7274         pos = find_next_bit(lrn->lrn_known_bitmap,
7275                             LFSCK_RBTREE_BITMAP_WIDTH, pos);
7276         if (pos >= LFSCK_RBTREE_BITMAP_WIDTH) {
7277                 key->f_oid = lrn->lrn_first_oid + pos;
7278                 if (unlikely(key->f_oid < lrn->lrn_first_oid)) {
7279                         key->f_seq++;
7280                         key->f_oid = 0;
7281                 }
7282                 it->loi_lrn = NULL;
7283                 goto again0;
7284         }
7285
7286         if (test_bit(pos, lrn->lrn_accessed_bitmap)) {
7287                 pos++;
7288                 goto again1;
7289         }
7290
7291         key->f_oid = lrn->lrn_first_oid + pos;
7292         obj = lfsck_object_find_bottom(env, lfsck, key);
7293         if (IS_ERR(obj)) {
7294                 rc = PTR_ERR(obj);
7295                 if (rc == -ENOENT) {
7296                         pos++;
7297                         goto again1;
7298                 }
7299                 RETURN(rc);
7300         }
7301
7302         dt_read_lock(env, obj, 0);
7303         if (dt_object_exists(obj) == 0 ||
7304             lfsck_is_dead_obj(obj)) {
7305                 dt_read_unlock(env, obj);
7306                 lfsck_object_put(env, obj);
7307                 pos++;
7308                 goto again1;
7309         }
7310
7311         rc = dt_attr_get(env, obj, la);
7312         if (rc != 0)
7313                 GOTO(out, rc);
7314
7315         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, ff, sizeof(*ff)),
7316                           XATTR_NAME_FID);
7317         if (rc == -ENODATA) {
7318                 /* For the pre-created OST-object, update the bitmap to avoid
7319                  * others LFSCK (second phase) iteration to touch it again. */
7320                 if (la->la_ctime == 0) {
7321                         if (!test_and_set_bit(pos, lrn->lrn_accessed_bitmap))
7322                                 atomic_inc(&lrn->lrn_accessed_count);
7323
7324                         /* For the race between repairing dangling referenced
7325                          * MDT-object and unlink the file, it may left orphan
7326                          * OST-object there. Destroy it now! */
7327                         if (unlikely(!(la->la_mode & S_ISUID))) {
7328                                 dt_read_unlock(env, obj);
7329                                 lfsck_layout_destroy_orphan(env, obj);
7330                                 lfsck_object_put(env, obj);
7331                                 pos++;
7332                                 goto again1;
7333                         }
7334                 } else if (idx == 0) {
7335                         /* If the orphan OST-object has no parent information,
7336                          * regard it as referenced by the MDT-object on MDT0. */
7337                         fid_zero(&rec->lor_rec.lor_fid);
7338                         rec->lor_rec.lor_uid = la->la_uid;
7339                         rec->lor_rec.lor_gid = la->la_gid;
7340                         memset(ol, 0, sizeof(*ol));
7341                         rec->lor_layout_version = 0;
7342                         rec->lor_range = 0;
7343
7344                         GOTO(out, rc = 0);
7345                 }
7346
7347                 dt_read_unlock(env, obj);
7348                 lfsck_object_put(env, obj);
7349                 pos++;
7350                 goto again1;
7351         }
7352
7353         if (rc < sizeof(struct lu_fid))
7354                 GOTO(out, rc = (rc < 0 ? rc : -EINVAL));
7355
7356         fid_le_to_cpu(&rec->lor_rec.lor_fid, &ff->ff_parent);
7357         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
7358          * MDT-object's FID::f_ver, instead it is the OST-object index in its
7359          * parent MDT-object's layout EA. */
7360         save = rec->lor_rec.lor_fid.f_stripe_idx;
7361         rec->lor_rec.lor_fid.f_ver = 0;
7362         rc = lfsck_fid_match_idx(env, lfsck, &rec->lor_rec.lor_fid, idx);
7363         /* If the orphan OST-object does not claim the MDT, then next.
7364          *
7365          * If we do not know whether it matches or not, then return it
7366          * to the MDT for further check. */
7367         if (rc == 0) {
7368                 dt_read_unlock(env, obj);
7369                 lfsck_object_put(env, obj);
7370                 pos++;
7371                 goto again1;
7372         }
7373
7374         rec->lor_rec.lor_fid.f_stripe_idx = save;
7375         rec->lor_rec.lor_uid = la->la_uid;
7376         rec->lor_rec.lor_gid = la->la_gid;
7377         ost_layout_le_to_cpu(ol, &ff->ff_layout);
7378         rec->lor_layout_version =
7379                 le32_to_cpu(ff->ff_layout_version & ~LU_LAYOUT_RESYNC);
7380         rec->lor_range = le32_to_cpu(ff->ff_range);
7381
7382         CDEBUG(D_LFSCK, "%s: return orphan "DFID", PFID "DFID", owner %u:%u, "
7383                "stripe size %u, stripe count %u, COMP id %u, COMP start %llu, "
7384                "COMP end %llu, layout version %u, range %u\n",
7385                lfsck_lfsck2name(com->lc_lfsck), PFID(key),
7386                PFID(&rec->lor_rec.lor_fid), rec->lor_rec.lor_uid,
7387                rec->lor_rec.lor_gid, ol->ol_stripe_size, ol->ol_stripe_count,
7388                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
7389                rec->lor_layout_version, rec->lor_range);
7390
7391         GOTO(out, rc = 0);
7392
7393 out:
7394         dt_read_unlock(env, obj);
7395         lfsck_object_put(env, obj);
7396         if (rc == 0)
7397                 it->loi_hash++;
7398
7399         return rc;
7400 }
7401
7402 /**
7403  * \retval       +1: locate to the exactly position
7404  * \retval        0: cannot locate to the exactly position,
7405  *                   call next() to move to a valid position.
7406  * \retval      -ve: on error
7407  */
7408 static int lfsck_orphan_it_get(const struct lu_env *env,
7409                                struct dt_it *di,
7410                                const struct dt_key *key)
7411 {
7412         struct lfsck_orphan_it  *it   = (struct lfsck_orphan_it *)di;
7413         int                      rc;
7414
7415         it->loi_key = *(struct lu_fid *)key;
7416         rc = lfsck_orphan_it_next(env, di);
7417         if (rc == 1)
7418                 return 0;
7419
7420         if (rc == 0)
7421                 return 1;
7422
7423         return rc;
7424 }
7425
7426 static void lfsck_orphan_it_put(const struct lu_env *env,
7427                                 struct dt_it *di)
7428 {
7429 }
7430
7431 static struct dt_key *lfsck_orphan_it_key(const struct lu_env *env,
7432                                           const struct dt_it *di)
7433 {
7434         struct lfsck_orphan_it *it = (struct lfsck_orphan_it *)di;
7435
7436         return (struct dt_key *)&it->loi_key;
7437 }
7438
7439 static int lfsck_orphan_it_key_size(const struct lu_env *env,
7440                                     const struct dt_it *di)
7441 {
7442         return sizeof(struct lu_fid);
7443 }
7444
7445 static int lfsck_orphan_it_rec(const struct lu_env *env,
7446                                const struct dt_it *di,
7447                                struct dt_rec *rec,
7448                                __u32 attr)
7449 {
7450         struct lfsck_orphan_it *it = (struct lfsck_orphan_it *)di;
7451
7452         *(struct lu_orphan_rec_v3 *)rec = it->loi_rec;
7453
7454         return 0;
7455 }
7456
7457 static __u64 lfsck_orphan_it_store(const struct lu_env *env,
7458                                    const struct dt_it *di)
7459 {
7460         struct lfsck_orphan_it  *it   = (struct lfsck_orphan_it *)di;
7461
7462         return it->loi_hash;
7463 }
7464
7465 /**
7466  * \retval       +1: locate to the exactly position
7467  * \retval        0: cannot locate to the exactly position,
7468  *                   call next() to move to a valid position.
7469  * \retval      -ve: on error
7470  */
7471 static int lfsck_orphan_it_load(const struct lu_env *env,
7472                                 const struct dt_it *di,
7473                                 __u64 hash)
7474 {
7475         struct lfsck_orphan_it           *it   = (struct lfsck_orphan_it *)di;
7476         struct lfsck_layout_slave_target *llst = it->loi_llst;
7477         int                               rc;
7478
7479         LASSERT(llst != NULL);
7480
7481         if (hash != llst->llst_hash) {
7482                 CDEBUG(D_LFSCK, "%s: the given hash %llu for orphan "
7483                        "iteration does not match the one when fini "
7484                        "%llu, to be reset.\n",
7485                        lfsck_lfsck2name(it->loi_com->lc_lfsck), hash,
7486                        llst->llst_hash);
7487                 fid_zero(&llst->llst_fid);
7488                 llst->llst_hash = 0;
7489         }
7490
7491         it->loi_key = llst->llst_fid;
7492         it->loi_hash = llst->llst_hash;
7493         rc = lfsck_orphan_it_next(env, (struct dt_it *)di);
7494         if (rc == 1)
7495                 return 0;
7496
7497         if (rc == 0)
7498                 return 1;
7499
7500         return rc;
7501 }
7502
7503 static int lfsck_orphan_it_key_rec(const struct lu_env *env,
7504                                    const struct dt_it *di,
7505                                    void *key_rec)
7506 {
7507         return 0;
7508 }
7509
7510 static const struct dt_index_operations lfsck_orphan_index_ops = {
7511         .dio_lookup             = lfsck_orphan_index_lookup,
7512         .dio_declare_insert     = lfsck_orphan_index_declare_insert,
7513         .dio_insert             = lfsck_orphan_index_insert,
7514         .dio_declare_delete     = lfsck_orphan_index_declare_delete,
7515         .dio_delete             = lfsck_orphan_index_delete,
7516         .dio_it = {
7517                 .init           = lfsck_orphan_it_init,
7518                 .fini           = lfsck_orphan_it_fini,
7519                 .get            = lfsck_orphan_it_get,
7520                 .put            = lfsck_orphan_it_put,
7521                 .next           = lfsck_orphan_it_next,
7522                 .key            = lfsck_orphan_it_key,
7523                 .key_size       = lfsck_orphan_it_key_size,
7524                 .rec            = lfsck_orphan_it_rec,
7525                 .store          = lfsck_orphan_it_store,
7526                 .load           = lfsck_orphan_it_load,
7527                 .key_rec        = lfsck_orphan_it_key_rec,
7528         }
7529 };